From mboxrd@z Thu Jan  1 00:00:00 1970
From: Alexander Aring <aahringo@redhat.com>
Date: Thu,  2 Dec 2021 15:20:39 -0500
Subject: [Cluster-devel] [PATCH dlm-tool 6/6] python: add ebpf histogram
	example
In-Reply-To: <20211202202039.1386193-1-aahringo@redhat.com>
References: <20211202202039.1386193-1-aahringo@redhat.com>
Message-ID: <20211202202039.1386193-7-aahringo@redhat.com>
List-Id: <cluster-devel.redhat.com>
To: cluster-devel.redhat.com
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit

This patch adds an example how a measurement between a kernel dlm_lock()
and their ast callback can be captured to get at the end a histogram.
This is hopefully a start of more useful dlm tracing collection with use
of ebpf.
---
 python/ebpf/README     |  6 ++++
 python/ebpf/dlmhist.py | 78 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 84 insertions(+)
 create mode 100644 python/ebpf/README
 create mode 100755 python/ebpf/dlmhist.py

diff --git a/python/ebpf/README b/python/ebpf/README
new file mode 100644
index 00000000..dec59ed0
--- /dev/null
+++ b/python/ebpf/README
@@ -0,0 +1,6 @@
+In this directory are some ebpf tracing examples. It requires bcc toolchain
+(usually known as bcc package name) and the python bcc module (usually known
+as python3-bcc package name).
+
+Also the current kernel headers need to be available or activate
+CONFIG_IKHEADERS in your kernel configuration.
diff --git a/python/ebpf/dlmhist.py b/python/ebpf/dlmhist.py
new file mode 100755
index 00000000..ed1eab4f
--- /dev/null
+++ b/python/ebpf/dlmhist.py
@@ -0,0 +1,78 @@
+#!/usr/bin/python
+#
+# This example shows how to capture latency between a dlm_lock() kernel
+# call for DLM_LOCK_EX requests with flag DLM_LKF_NOQUEUE and the ast
+# response.
+#
+# You will probably see two line peaks, one in case of that the current
+# node is the lock master and another one which requires network
+# communication. There is currently no way to filter them out, so the
+# second peak is interessting to get an idea what time it takes to
+# call dlm_lock() and get a response back.
+
+from bcc import BPF
+
+import threading
+
+b = BPF(text="""
+#include <uapi/linux/ptrace.h>
+#include <uapi/linux/dlm.h>
+
+BPF_HASH(start, u32);
+BPF_HISTOGRAM(dist, u64);
+
+#define DLM_HASH(args) (args->ls_id ^ args->lkb_id)
+
+TRACEPOINT_PROBE(dlm, dlm_lock_start)
+{
+    u64 ts = bpf_ktime_get_ns();
+    u32 hash = DLM_HASH(args);
+
+    if (args->flags & DLM_LKF_NOQUEUE &&
+        args->mode == DLM_LOCK_EX)
+        start.update(&hash, &ts);
+
+    return 0;
+}
+
+TRACEPOINT_PROBE(dlm, dlm_lock_end)
+{
+    u32 hash = DLM_HASH(args);
+
+    if (args->error != 0)
+        start.delete(&hash);
+
+    return 0;
+}
+
+TRACEPOINT_PROBE(dlm, dlm_ast)
+{
+    u32 hash = DLM_HASH(args);
+    u64 *tsp, delta;
+
+    tsp = start.lookup(&hash);
+    if (tsp != 0) {
+        start.delete(&hash);
+        delta = bpf_ktime_get_ns() - *tsp;
+
+        if (args->sb_status != 0)
+            return 0;
+
+        dist.increment(bpf_log2l(delta));
+    }
+
+    return 0;
+}
+""")
+
+print("Tracing... Hit Ctrl-C anytime to end.")
+
+forever = threading.Event()
+try:
+    forever.wait()
+except KeyboardInterrupt:
+    print()
+
+print("log2 histogram")
+print("--------------")
+b["dist"].print_log2_hist("ns")
-- 
2.27.0