mdflush

README.md

@@ -85,6 +85,7 @@ Tools:
 - tools/[gethostlatency](tools/gethostlatency.py): Show latency for getaddrinfo/gethostbyname[2] calls. [Examples](tools/gethostlatency_example.txt).
 - tools/[hardirqs](tools/hardirqs.py):  Measure hard IRQ (hard interrupt) event time. [Examples](tools/hardirqs_example.txt).
 - tools/[killsnoop](tools/killsnoop.py): Trace signals issued by the kill() syscall. [Examples](tools/killsnoop_example.txt).
+- tools/[mdflush](tools/mdflush.py): Trace md flush events. [Examples](tools/mdflush.txt).
 - tools/[memleak](tools/memleak.py): Display outstanding memory allocations to find memory leaks. [Examples](tools/memleak_examples.txt).
 - tools/[offcputime](tools/offcputime.py): Summarize off-CPU time by kernel stack trace. [Examples](tools/offcputime_example.txt).
 - tools/[offwaketime](tools/offwaketime.py): Summarize blocked time by kernel off-CPU stack and waker stack. [Examples](tools/offwaketime_example.txt).

man/man8/mdflush.8

@@ -0,0 +1,57 @@
+.TH pidpersec 8  "2016-02-13" "USER COMMANDS"
+.SH NAME
+mdflush \- Trace md flush events. Uses Linux eBPF/bcc.
+.SH SYNOPSIS
+.B mdflush
+.SH DESCRIPTION
+This tool traces flush events by md, the Linux multiple device driver
+(software RAID). The timestamp and md device for the flush are printed.
+Knowing when these flushes happen can be useful for correlation with
+unexplained spikes in disk latency.
+
+This works by tracing the kernel md_flush_request() function using dynamic
+tracing, and will need updating to match any changes to this function.
+
+Note that the flushes themselves are likely to originate from higher in the
+I/O stack, such as from the file systems.
+
+Since this uses BPF, only the root user can use this tool.
+.SH REQUIREMENTS
+CONFIG_BPF and bcc.
+.SH EXAMPLES
+.TP
+Trace md flush events:
+#
+.B mdflush
+.SH FIELDS
+.TP
+TIME
+Time of the flush event (HH:MM:SS).
+.TP
+PID
+The process ID that was on-CPU when the event was issued. This may identify
+the cause of the flush (eg, the "sync" command), but will often identify a
+kernel worker thread that was managing I/O.
+.TP
+COMM
+The command name for the PID.
+.TP
+DEVICE
+The md device name.
+.SH OVERHEAD
+Expected to be negligible.
+.SH SOURCE
+This is from bcc.
+.IP
+https://github.com/iovisor/bcc
+.PP
+Also look in the bcc distribution for a companion _examples.txt file containing
+example usage, output, and commentary for this tool.
+.SH OS
+Linux
+.SH STABILITY
+Unstable - in development.
+.SH AUTHOR
+Brendan Gregg
+.SH SEE ALSO
+biosnoop(8)

tools/mdflush.py

@@ -0,0 +1,68 @@
+#!/usr/bin/python
+# @lint-avoid-python-3-compatibility-imports
+#
+# mdflush  Trace md flush events.
+#          For Linux, uses BCC, eBPF.
+#
+# Todo: add more details of the flush (latency, I/O count).
+#
+# Copyright 2016 Netflix, Inc.
+# Licensed under the Apache License, Version 2.0 (the "License")
+#
+# 13-Feb-2015   Brendan Gregg   Created this.
+
+from __future__ import print_function
+from bcc import BPF
+from time import strftime
+import ctypes as ct
+
+# load BPF program
+b = BPF(text="""
+#include <uapi/linux/ptrace.h>
+#include <linux/sched.h>
+#include <linux/genhd.h>
+
+struct data_t {
+    u64 pid;
+    char comm[TASK_COMM_LEN];
+    char disk[DISK_NAME_LEN];
+};
+BPF_PERF_OUTPUT(events);
+
+int kprobe__md_flush_request(struct pt_regs *ctx, void *mddev, struct bio *bio)
+{
+    struct data_t data = {};
+    u32 pid = bpf_get_current_pid_tgid();
+    data.pid = pid;
+    bpf_get_current_comm(&data.comm, sizeof(data.comm));
+    bpf_probe_read(&data.disk, sizeof(data.disk),
+        bio->bi_bdev->bd_disk->disk_name);
+    events.perf_submit(ctx, &data, sizeof(data));
+    return 0;
+}
+""")
+
+# event data
+TASK_COMM_LEN = 16  # linux/sched.h
+DISK_NAME_LEN = 32  # linux/genhd.h
+class Data(ct.Structure):
+    _fields_ = [
+        ("pid", ct.c_ulonglong),
+        ("comm", ct.c_char * TASK_COMM_LEN),
+        ("disk", ct.c_char * DISK_NAME_LEN)
+    ]
+
+# header
+print("Tracing md flush requests... Hit Ctrl-C to end.")
+print("%-8s %-6s %-16s %s" % ("TIME", "PID", "COMM", "DEVICE"))
+
+# process event
+def print_event(cpu, data, size):
+    event = ct.cast(data, ct.POINTER(Data)).contents
+    print("%-8s %-6d %-16s %s" % (strftime("%H:%M:%S"), event.pid, event.comm,
+        event.disk))
+
+# read events
+b["events"].open_perf_buffer(print_event)
+while 1:
+    b.kprobe_poll()

tools/mdflush_example.txt

@@ -0,0 +1,47 @@
+Demonstrations of mdflush, the Linux eBPF/bcc version.
+
+
+The mdflush tool traces flushes at the md driver level, and prints details
+including the time of the flush:
+
+# ./mdflush
+Tracing md flush requests... Hit Ctrl-C to end.
+TIME     PID    COMM             DEVICE
+03:13:49 16770  sync             md0
+03:14:08 16864  sync             md0
+03:14:49 496    kworker/1:0H     md0
+03:14:49 488    xfsaild/md0      md0
+03:14:54 488    xfsaild/md0      md0
+03:15:00 488    xfsaild/md0      md0
+03:15:02 85     kswapd0          md0
+03:15:02 488    xfsaild/md0      md0
+03:15:05 488    xfsaild/md0      md0
+03:15:08 488    xfsaild/md0      md0
+03:15:10 488    xfsaild/md0      md0
+03:15:11 488    xfsaild/md0      md0
+03:15:11 488    xfsaild/md0      md0
+03:15:11 488    xfsaild/md0      md0
+03:15:11 488    xfsaild/md0      md0
+03:15:11 488    xfsaild/md0      md0
+03:15:12 488    xfsaild/md0      md0
+03:15:13 488    xfsaild/md0      md0
+03:15:15 488    xfsaild/md0      md0
+03:15:19 496    kworker/1:0H     md0
+03:15:49 496    kworker/1:0H     md0
+03:15:55 18840  sync             md0
+03:16:49 496    kworker/1:0H     md0
+03:17:19 496    kworker/1:0H     md0
+03:20:19 496    kworker/1:0H     md0
+03:21:19 496    kworker/1:0H     md0
+03:21:49 496    kworker/1:0H     md0
+03:25:19 496    kworker/1:0H     md0
+[...]
+
+This can be useful for correlation with latency outliers or spikes in disk
+latency, as measured using another tool (eg, system monitoring). If spikes in
+disk latency often coincide with md flush events, then it would make flushing
+a target for tuning.
+
+Note that the flush events are likely to originate from higher in the I/O
+stack, such as from file systems. This traces md processing them, and the
+timestamp corresponds with when md began to issue the flush to disks.