bpf: BPF based latency tracing

author Daniel Wagner <daniel.wagner@bmw-carit.de>

Fri, 19 Jun 2015 14:00:44 +0000 (16:00 +0200)

committer David S. Miller <davem@davemloft.net>

Tue, 23 Jun 2015 13:09:58 +0000 (06:09 -0700)
author Daniel Wagner <daniel.wagner@bmw-carit.de>
Fri, 19 Jun 2015 14:00:44 +0000 (16:00 +0200)
committer David S. Miller <davem@davemloft.net>
Tue, 23 Jun 2015 13:09:58 +0000 (06:09 -0700)
diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile

index 46c6a8cf74d3e56cad920616208b9d603a08153c..4450fed91ab40f61306ab1af8bf38d470b1f4114 100644 (file)
--- a/samples/bpf/Makefile
+++ b/samples/bpf/Makefile
@@ -12,6 +12,7 @@ hostprogs-y += tracex2
  hostprogs-y += tracex3
  hostprogs-y += tracex4
  hostprogs-y += tracex5
+hostprogs-y += lathist
  
  test_verifier-objs := test_verifier.o libbpf.o
  test_maps-objs := test_maps.o libbpf.o
@@ -24,6 +25,7 @@ tracex2-objs := bpf_load.o libbpf.o tracex2_user.o
  tracex3-objs := bpf_load.o libbpf.o tracex3_user.o
  tracex4-objs := bpf_load.o libbpf.o tracex4_user.o
  tracex5-objs := bpf_load.o libbpf.o tracex5_user.o
+lathist-objs := bpf_load.o libbpf.o lathist_user.o
  
  # Tell kbuild to always build the programs
  always := $(hostprogs-y)
@@ -36,6 +38,7 @@ always += tracex3_kern.o
  always += tracex4_kern.o
  always += tracex5_kern.o
  always += tcbpf1_kern.o
+always += lathist_kern.o
  
  HOSTCFLAGS += -I$(objtree)/usr/include
  
@@ -48,6 +51,7 @@ HOSTLOADLIBES_tracex2 += -lelf
  HOSTLOADLIBES_tracex3 += -lelf
  HOSTLOADLIBES_tracex4 += -lelf -lrt
  HOSTLOADLIBES_tracex5 += -lelf
+HOSTLOADLIBES_lathist += -lelf
  
  # point this to your LLVM backend with bpf support
  LLC=$(srctree)/tools/bpf/llvm/bld/Debug+Asserts/bin/llc
diff --git a/samples/bpf/lathist_kern.c b/samples/bpf/lathist_kern.c

new file mode 100644 (file)

index 0000000..18fa088
--- /dev/null
+++ b/samples/bpf/lathist_kern.c
@@ -0,0 +1,99 @@
+/* Copyright (c) 2013-2015 PLUMgrid, http://plumgrid.com
+ * Copyright (c) 2015 BMW Car IT GmbH
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ */
+#include <linux/version.h>
+#include <linux/ptrace.h>
+#include <uapi/linux/bpf.h>
+#include "bpf_helpers.h"
+
+#define MAX_ENTRIES    20
+#define MAX_CPU                4
+
+/* We need to stick to static allocated memory (an array instead of
+ * hash table) because managing dynamic memory from the
+ * trace_preempt_[on|off] tracepoints hooks is not supported.
+ */
+
+struct bpf_map_def SEC("maps") my_map = {
+       .type = BPF_MAP_TYPE_ARRAY,
+       .key_size = sizeof(int),
+       .value_size = sizeof(u64),
+       .max_entries = MAX_CPU,
+};
+
+SEC("kprobe/trace_preempt_off")
+int bpf_prog1(struct pt_regs *ctx)
+{
+       int cpu = bpf_get_smp_processor_id();
+       u64 *ts = bpf_map_lookup_elem(&my_map, &cpu);
+
+       if (ts)
+               *ts = bpf_ktime_get_ns();
+
+       return 0;
+}
+
+static unsigned int log2(unsigned int v)
+{
+       unsigned int r;
+       unsigned int shift;
+
+       r = (v > 0xFFFF) << 4; v >>= r;
+       shift = (v > 0xFF) << 3; v >>= shift; r |= shift;
+       shift = (v > 0xF) << 2; v >>= shift; r |= shift;
+       shift = (v > 0x3) << 1; v >>= shift; r |= shift;
+       r |= (v >> 1);
+
+       return r;
+}
+
+static unsigned int log2l(unsigned long v)
+{
+       unsigned int hi = v >> 32;
+
+       if (hi)
+               return log2(hi) + 32;
+       else
+               return log2(v);
+}
+
+struct bpf_map_def SEC("maps") my_lat = {
+       .type = BPF_MAP_TYPE_ARRAY,
+       .key_size = sizeof(int),
+       .value_size = sizeof(long),
+       .max_entries = MAX_CPU * MAX_ENTRIES,
+};
+
+SEC("kprobe/trace_preempt_on")
+int bpf_prog2(struct pt_regs *ctx)
+{
+       u64 *ts, cur_ts, delta;
+       int key, cpu;
+       long *val;
+
+       cpu = bpf_get_smp_processor_id();
+       ts = bpf_map_lookup_elem(&my_map, &cpu);
+       if (!ts)
+               return 0;
+
+       cur_ts = bpf_ktime_get_ns();
+       delta = log2l(cur_ts - *ts);
+
+       if (delta > MAX_ENTRIES - 1)
+               delta = MAX_ENTRIES - 1;
+
+       key = cpu * MAX_ENTRIES + delta;
+       val = bpf_map_lookup_elem(&my_lat, &key);
+       if (val)
+               __sync_fetch_and_add((long *)val, 1);
+
+       return 0;
+
+}
+
+char _license[] SEC("license") = "GPL";
+u32 _version SEC("version") = LINUX_VERSION_CODE;
diff --git a/samples/bpf/lathist_user.c b/samples/bpf/lathist_user.c

new file mode 100644 (file)

index 0000000..65da8c1
--- /dev/null
+++ b/samples/bpf/lathist_user.c
@@ -0,0 +1,103 @@
+/* Copyright (c) 2013-2015 PLUMgrid, http://plumgrid.com
+ * Copyright (c) 2015 BMW Car IT GmbH
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ */
+#include <stdio.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <signal.h>
+#include <linux/bpf.h>
+#include "libbpf.h"
+#include "bpf_load.h"
+
+#define MAX_ENTRIES    20
+#define MAX_CPU                4
+#define MAX_STARS      40
+
+struct cpu_hist {
+       long data[MAX_ENTRIES];
+       long max;
+};
+
+static struct cpu_hist cpu_hist[MAX_CPU];
+
+static void stars(char *str, long val, long max, int width)
+{
+       int i;
+
+       for (i = 0; i < (width * val / max) - 1 && i < width - 1; i++)
+               str[i] = '*';
+       if (val > max)
+               str[i - 1] = '+';
+       str[i] = '\0';
+}
+
+static void print_hist(void)
+{
+       char starstr[MAX_STARS];
+       struct cpu_hist *hist;
+       int i, j;
+
+       /* clear screen */
+       printf("\033[2J");
+
+       for (j = 0; j < MAX_CPU; j++) {
+               hist = &cpu_hist[j];
+
+               /* ignore CPUs without data (maybe offline?) */
+               if (hist->max == 0)
+                       continue;
+
+               printf("CPU %d\n", j);
+               printf("      latency        : count     distribution\n");
+               for (i = 1; i <= MAX_ENTRIES; i++) {
+                       stars(starstr, hist->data[i - 1], hist->max, MAX_STARS);
+                       printf("%8ld -> %-8ld : %-8ld |%-*s|\n",
+                               (1l << i) >> 1, (1l << i) - 1,
+                               hist->data[i - 1], MAX_STARS, starstr);
+               }
+       }
+}
+
+static void get_data(int fd)
+{
+       long key, value;
+       int c, i;
+
+       for (i = 0; i < MAX_CPU; i++)
+               cpu_hist[i].max = 0;
+
+       for (c = 0; c < MAX_CPU; c++) {
+               for (i = 0; i < MAX_ENTRIES; i++) {
+                       key = c * MAX_ENTRIES + i;
+                       bpf_lookup_elem(fd, &key, &value);
+
+                       cpu_hist[c].data[i] = value;
+                       if (value > cpu_hist[c].max)
+                               cpu_hist[c].max = value;
+               }
+       }
+}
+
+int main(int argc, char **argv)
+{
+       char filename[256];
+
+       snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
+
+       if (load_bpf_file(filename)) {
+               printf("%s", bpf_log_buf);
+               return 1;
+       }
+
+       while (1) {
+               get_data(map_fd[1]);
+               print_hist();
+               sleep(5);
+       }
+
+       return 0;
+}
author	Daniel Wagner <daniel.wagner@bmw-carit.de>
	Fri, 19 Jun 2015 14:00:44 +0000 (16:00 +0200)
committer	David S. Miller <davem@davemloft.net>
	Tue, 23 Jun 2015 13:09:58 +0000 (06:09 -0700)
samples/bpf/Makefile		patch \| blob \| history
samples/bpf/lathist_kern.c	[new file with mode: 0644]	patch \| blob
samples/bpf/lathist_user.c	[new file with mode: 0644]	patch \| blob