On 3/6/20 4:17 PM, Song Liu wrote:
With fentry/fexit programs, it is possible to profile BPF program with
hardware counters. Introduce bpftool "prog profile", which measures key
metrics of a BPF program.
bpftool prog profile command creates per-cpu perf events. Then it attaches
fentry/fexit programs to the target BPF program. The fentry program saves
perf event value to a map. The fexit program reads the perf event again,
and calculates the difference, which is the instructions/cycles used by
the target program.
Example input and output:
./bpftool prog profile id 337 duration 3 cycles instructions llc_misses
4228 run_cnt
3403698 cycles (84.08%)
3525294 instructions # 1.04 insn per cycle (84.05%)
13 llc_misses # 3.69 LLC misses per million isns (83.50%)
This command measures cycles and instructions for BPF program with id
337 for 3 seconds. The program has triggered 4228 times. The rest of the
output is similar to perf-stat. In this example, the counters were only
counting ~84% of the time because of time multiplexing of perf counters.
Note that, this approach measures cycles and instructions in very small
increments. So the fentry/fexit programs introduce noticeable errors to
the measurement results.
The fentry/fexit programs are generated with BPF skeletons. Therefore, we
build bpftool twice. The first time _bpftool is built without skeletons.
Then, _bpftool is used to generate the skeletons. The second time, bpftool
is built with skeletons.
Reviewed-by: Quentin Monnet <quentin@xxxxxxxxxxxxx>
Signed-off-by: Song Liu <songliubraving@xxxxxx>
---
tools/bpf/bpftool/Makefile | 18 +
tools/bpf/bpftool/prog.c | 424 +++++++++++++++++++++-
tools/bpf/bpftool/skeleton/profiler.bpf.c | 171 +++++++++
tools/bpf/bpftool/skeleton/profiler.h | 47 +++
tools/scripts/Makefile.include | 1 +
5 files changed, 660 insertions(+), 1 deletion(-)
create mode 100644 tools/bpf/bpftool/skeleton/profiler.bpf.c
create mode 100644 tools/bpf/bpftool/skeleton/profiler.h
[...]
diff --git a/tools/bpf/bpftool/skeleton/profiler.bpf.c b/tools/bpf/bpftool/skeleton/profiler.bpf.c
new file mode 100644
index 000000000000..20594ccb393d
--- /dev/null
+++ b/tools/bpf/bpftool/skeleton/profiler.bpf.c
@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: GPL-2.0
+// Copyright (c) 2020 Facebook
+#include "profiler.h"
+#include <linux/bpf.h>
+#include <bpf/bpf_helpers.h>
+
+#define ___bpf_concat(a, b) a ## b
+#define ___bpf_apply(fn, n) ___bpf_concat(fn, n)
+#define ___bpf_nth(_, _1, _2, _3, _4, _5, _6, _7, _8, _9, _a, _b, _c, N, ...) N
+#define ___bpf_narg(...) \
+ ___bpf_nth(_, ##__VA_ARGS__, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
+#define ___bpf_empty(...) \
+ ___bpf_nth(_, ##__VA_ARGS__, N, N, N, N, N, N, N, N, N, N, 0)
+
+#define ___bpf_ctx_cast0() ctx
+#define ___bpf_ctx_cast1(x) ___bpf_ctx_cast0(), (void *)ctx[0]
+#define ___bpf_ctx_cast2(x, args...) ___bpf_ctx_cast1(args), (void *)ctx[1]
+#define ___bpf_ctx_cast3(x, args...) ___bpf_ctx_cast2(args), (void *)ctx[2]
+#define ___bpf_ctx_cast4(x, args...) ___bpf_ctx_cast3(args), (void *)ctx[3]
+#define ___bpf_ctx_cast5(x, args...) ___bpf_ctx_cast4(args), (void *)ctx[4]
+#define ___bpf_ctx_cast6(x, args...) ___bpf_ctx_cast5(args), (void *)ctx[5]
+#define ___bpf_ctx_cast7(x, args...) ___bpf_ctx_cast6(args), (void *)ctx[6]
+#define ___bpf_ctx_cast8(x, args...) ___bpf_ctx_cast7(args), (void *)ctx[7]
+#define ___bpf_ctx_cast9(x, args...) ___bpf_ctx_cast8(args), (void *)ctx[8]
+#define ___bpf_ctx_cast10(x, args...) ___bpf_ctx_cast9(args), (void *)ctx[9]
+#define ___bpf_ctx_cast11(x, args...) ___bpf_ctx_cast10(args), (void *)ctx[10]
+#define ___bpf_ctx_cast12(x, args...) ___bpf_ctx_cast11(args), (void *)ctx[11]
+#define ___bpf_ctx_cast(args...) \
+ ___bpf_apply(___bpf_ctx_cast, ___bpf_narg(args))(args)
+
+/*
+ * BPF_PROG is a convenience wrapper for generic tp_btf/fentry/fexit and
+ * similar kinds of BPF programs, that accept input arguments as a single
+ * pointer to untyped u64 array, where each u64 can actually be a typed
+ * pointer or integer of different size. Instead of requring user to write
+ * manual casts and work with array elements by index, BPF_PROG macro
+ * allows user to declare a list of named and typed input arguments in the
+ * same syntax as for normal C function. All the casting is hidden and
+ * performed transparently, while user code can just assume working with
+ * function arguments of specified type and name.
+ *
+ * Original raw context argument is preserved as well as 'ctx' argument.
+ * This is useful when using BPF helpers that expect original context
+ * as one of the parameters (e.g., for bpf_perf_event_output()).
+ */
+#define BPF_PROG(name, args...) \
+name(unsigned long long *ctx); \
+static __always_inline typeof(name(0)) \
+____##name(unsigned long long *ctx, ##args); \
+typeof(name(0)) name(unsigned long long *ctx) \
+{ \
+ _Pragma("GCC diagnostic push") \
+ _Pragma("GCC diagnostic ignored \"-Wint-conversion\"") \
+ return ____##name(___bpf_ctx_cast(args)); \
+ _Pragma("GCC diagnostic pop") \
+} \
+static __always_inline typeof(name(0)) \
+____##name(unsigned long long *ctx, ##args)
The above change has merged into tools/lib/bpf/bpf_tracing.h.
You can remove them by just including "bpf/bpf_tracing.h"?
