migrated gethostlatency to use bpf_perf_event

man/man8/gethostlatency.8

@@ -11,6 +11,10 @@ latency of the call (duration) in milliseconds, and the host string.
 This tool can be useful for identifying DNS latency, by identifying which
 remote host name lookups were slow, and by how much.
 
+This makes use of a Linux 4.5 feature (bpf_perf_event_output());
+for kernels older than 4.5, see the version under tools/old,
+which uses an older mechanism
+
 This tool currently uses dynamic tracing of user-level functions and registers,
 and may need modifications to match your software and processor architecture.
 

tools/gethostlatency.py

@@ -13,43 +13,71 @@
 # Licensed under the Apache License, Version 2.0 (the "License")
 #
 # 28-Jan-2016 Brendan Gregg Created this.
+# 30-Mar-2016 Allan McAleavy updated for BPF_PERF_OUTPUT
 
 from __future__ import print_function
 from bcc import BPF
 from time import strftime
+import ctypes as ct
 
 # load BPF program
 bpf_text = """
 #include <uapi/linux/ptrace.h>
+#include <linux/sched.h>
 
 struct val_t {
+ u32 pid;
+ char comm[TASK_COMM_LEN];
  char host[80];
  u64 ts;
 };
+
+struct data_t {
+ u32 pid;
+ u64 ts;
+ u64 delta;
+ char comm[TASK_COMM_LEN];
+ char host[80];
+};
+
 BPF_HASH(start, u32, struct val_t);
+BPF_PERF_OUTPUT(events);
 
 int do_entry(struct pt_regs *ctx) {
  if (!ctx->di)
  return 0;
+
  struct val_t val = {};
  u32 pid = bpf_get_current_pid_tgid();
- bpf_probe_read(&val.host, sizeof(val.host), (void *)ctx->di);
- val.ts = bpf_ktime_get_ns();
- start.update(&pid, &val);
+
+ if (bpf_get_current_comm(&val.comm, sizeof(val.comm)) == 0) {
+ bpf_probe_read(&val.host, sizeof(val.host), (void *)ctx->di);
+ val.pid = bpf_get_current_pid_tgid();
+ val.ts = bpf_ktime_get_ns();
+ start.update(&pid, &val);
+ }
+
  return 0;
 }
 
 int do_return(struct pt_regs *ctx) {
  struct val_t *valp;
+ struct data_t data = {};
  u64 delta;
  u32 pid = bpf_get_current_pid_tgid();
 
+ u64 tsp = bpf_ktime_get_ns();
+
  valp = start.lookup(&pid);
  if (valp == 0)
  return 0; // missed start
 
- delta = (bpf_ktime_get_ns() - valp->ts) / 1000;
- bpf_trace_printk("%d %s\\n", delta, valp->host);
+ bpf_probe_read(&data.comm, sizeof(data.comm), valp->comm);
+ bpf_probe_read(&data.host, sizeof(data.host), (void *)valp->host);
+ data.pid = valp->pid;
+ data.delta = tsp - valp->ts;
+ data.ts = tsp / 1000;
+ events.perf_submit(ctx, &data, sizeof(data));
  start.delete(&pid);
  return 0;
 }
@@ -62,16 +90,43 @@
 b.attach_uretprobe(name="c", sym="gethostbyname", fn_name="do_return")
 b.attach_uretprobe(name="c", sym="gethostbyname2", fn_name="do_return")
 
+TASK_COMM_LEN = 16 # linux/sched.h
+
+class Data(ct.Structure):
+ _fields_ = [
+ ("pid", ct.c_ulonglong),
+ ("ts", ct.c_ulonglong),
+ ("delta", ct.c_ulonglong),
+ ("comm", ct.c_char * TASK_COMM_LEN),
+ ("host", ct.c_char * 80)
+ ]
+
+start_ts = 0
+prev_ts = 0
+delta = 0
+
 # header
-print("%-9s %-6s %-12s %6s %s" % ("TIME", "PID", "COMM", "LATms", "HOST"))
+print("%-9s %-6s %-16s %10s %s" % ("TIME", "PID", "COMM", "LATms", "HOST"))
+
+def print_event(cpu, data, size):
+ event = ct.cast(data, ct.POINTER(Data)).contents
+ global start_ts
+ global prev_ts
+ global delta
+
+ if start_ts == 0:
+ prev_ts = start_ts
+
+ if start_ts == 1:
+ delta = float(delta) + (event.ts - prev_ts)
+
+ print("%-9s %-6d %-16s %10.2f %s" % (strftime("%H:%M:%S"), event.pid,
+ event.comm, (event.delta / 1000000), event.host))
+
+ prev_ts = event.ts
+ start_ts = 1
 
-# format output
+# loop with callback to print_event
+b["events"].open_perf_buffer(print_event)
 while 1:
- try:
- (task, pid, cpu, flags, ts, msg) = b.trace_fields()
- except ValueError:
- continue
- (delta, host) = msg.split(" ")
- deltams = int(delta) / 1000
- print("%-9s %-6d %-12.12s %6.2f %s" % (strftime("%H:%M:%S"), pid, task,
- deltams, host))
+ b.kprobe_poll()

tools/old/gethostlatency.py

@@ -0,0 +1,77 @@
+#!/usr/bin/python
+#
+# gethostlatency Show latency for getaddrinfo/gethostbyname[2] calls.
+# For Linux, uses BCC, eBPF. Embedded C.
+#
+# This can be useful for identifying DNS latency, by identifying which
+# remote host name lookups were slow, and by how much.
+#
+# This uses dynamic tracing of user-level functions and registers, and may
+# need modifications to match your software and processor architecture.
+#
+# Copyright 2016 Netflix, Inc.
+# Licensed under the Apache License, Version 2.0 (the "License")
+#
+# 28-Jan-2016 Brendan Gregg Created this.
+
+from __future__ import print_function
+from bcc import BPF
+from time import strftime
+
+# load BPF program
+bpf_text = """
+#include <uapi/linux/ptrace.h>
+
+struct val_t {
+ char host[80];
+ u64 ts;
+};
+BPF_HASH(start, u32, struct val_t);
+
+int do_entry(struct pt_regs *ctx) {
+ if (!ctx->di)
+ return 0;
+ struct val_t val = {};
+ u32 pid = bpf_get_current_pid_tgid();
+ bpf_probe_read(&val.host, sizeof(val.host), (void *)ctx->di);
+ val.ts = bpf_ktime_get_ns();
+ start.update(&pid, &val);
+ return 0;
+}
+
+int do_return(struct pt_regs *ctx) {
+ struct val_t *valp;
+ u64 delta;
+ u32 pid = bpf_get_current_pid_tgid();
+
+ valp = start.lookup(&pid);
+ if (valp == 0)
+ return 0; // missed start
+
+ delta = (bpf_ktime_get_ns() - valp->ts) / 1000;
+ bpf_trace_printk("%d %s\\n", delta, valp->host);
+ start.delete(&pid);
+ return 0;
+}
+"""
+b = BPF(text=bpf_text)
+b.attach_uprobe(name="c", sym="getaddrinfo", fn_name="do_entry")
+b.attach_uprobe(name="c", sym="gethostbyname", fn_name="do_entry")
+b.attach_uprobe(name="c", sym="gethostbyname2", fn_name="do_entry")
+b.attach_uretprobe(name="c", sym="getaddrinfo", fn_name="do_return")
+b.attach_uretprobe(name="c", sym="gethostbyname", fn_name="do_return")
+b.attach_uretprobe(name="c", sym="gethostbyname2", fn_name="do_return")
+
+# header
+print("%-9s %-6s %-12s %6s %s" % ("TIME", "PID", "COMM", "LATms", "HOST"))
+
+# format output
+while 1:
+ try:
+ (task, pid, cpu, flags, ts, msg) = b.trace_fields()
+ except ValueError:
+ continue
+ (delta, host) = msg.split(" ")
+ deltams = int(delta) / 1000
+ print("%-9s %-6d %-12.12s %6.2f %s" % (strftime("%H:%M:%S"), pid, task,
+ deltams, host))