Message-ID: <87lfp6dtvw.fsf@toke.dk>
Date:   Thu, 13 Feb 2020 23:06:43 +0100
From:   Toke Høiland-Jørgensen <toke@...hat.com>
To:     Song Liu <songliubraving@...com>
Cc:     Networking <netdev@...r.kernel.org>,
        "bpf\@vger.kernel.org" <bpf@...r.kernel.org>,
        Kernel Team <Kernel-team@...com>,
        "ast\@kernel.org" <ast@...nel.org>,
        "daniel\@iogearbox.net" <daniel@...earbox.net>
Subject: Re: [RFC bpf-next 3/4] bpftool: introduce "prog profile" command

Song Liu <songliubraving@...com> writes:

>> On Feb 13, 2020, at 1:50 PM, Toke Høiland-Jørgensen <toke@...hat.com> wrote:
>> 
>> Song Liu <songliubraving@...com> writes:
>> 
>>> 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 20 id 810 cycles instructions
>>>  cycles: duration 20 run_cnt 1368 miss_cnt 665
>>>          counter 503377 enabled 668202 running 351857
>>>  instructions: duration 20 run_cnt 1368 miss_cnt 707
>>>          counter 398625 enabled 502330 running 272014
>>> 
>>> This command measures cycles and instructions for BPF program with id
>>> 810 for 20 seconds. The program has triggered 1368 times. cycles was not
>>> measured in 665 out of these runs, because of perf event multiplexing
>>> (some perf commands are running in the background). In these runs, the BPF
>>> program consumed 503377 cycles. The perf_event enabled and running time
>>> are 668202 and 351857 respectively.
>>> 
>>> Note that, this approach measures cycles and instructions in very small
>>> increments. So the fentry/fexit programs introduce noticable errors to
>>> the measurement results.
>>> 
>>> The fentry/fexit programs are generated with BPF skeleton. Currently,
>>> generation of the skeleton requires some manual steps.
>>> 
>>> Signed-off-by: Song Liu <songliubraving@...com>
>>> ---
>>> tools/bpf/bpftool/profiler.skel.h         | 820 ++++++++++++++++++++++
>>> tools/bpf/bpftool/prog.c                  | 387 +++++++++-
>>> tools/bpf/bpftool/skeleton/README         |   3 +
>>> tools/bpf/bpftool/skeleton/profiler.bpf.c | 185 +++++
>>> tools/bpf/bpftool/skeleton/profiler.h     |  47 ++
>>> 5 files changed, 1441 insertions(+), 1 deletion(-)
>>> create mode 100644 tools/bpf/bpftool/profiler.skel.h
>>> create mode 100644 tools/bpf/bpftool/skeleton/README
>>> 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/profiler.skel.h b/tools/bpf/bpftool/profiler.skel.h
>>> new file mode 100644
>>> index 000000000000..10e99989c03e
>>> --- /dev/null
>>> +++ b/tools/bpf/bpftool/profiler.skel.h
>>> @@ -0,0 +1,820 @@
>>> +/* SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) */
>>> +
>>> +/* THIS FILE IS AUTOGENERATED! */
>>> +#ifndef __PROFILER_BPF_SKEL_H__
>>> +#define __PROFILER_BPF_SKEL_H__
>>> +
>>> +#include <stdlib.h>
>>> +#include <bpf/libbpf.h>
>>> +
>>> +struct profiler_bpf {
>>> +	struct bpf_object_skeleton *skeleton;
>>> +	struct bpf_object *obj;
>>> +	struct {
>>> +		struct bpf_map *events;
>>> +		struct bpf_map *fentry_readings;
>>> +		struct bpf_map *accum_readings;
>>> +		struct bpf_map *counts;
>>> +		struct bpf_map *miss_counts;
>>> +		struct bpf_map *rodata;
>>> +	} maps;
>>> +	struct {
>>> +		struct bpf_program *fentry_XXX;
>>> +		struct bpf_program *fexit_XXX;
>>> +	} progs;
>>> +	struct {
>>> +		struct bpf_link *fentry_XXX;
>>> +		struct bpf_link *fexit_XXX;
>>> +	} links;
>>> +	struct profiler_bpf__rodata {
>>> +		__u32 num_cpu;
>>> +		__u32 num_metric;
>>> +	} *rodata;
>>> +};
>>> +
>>> +static void
>>> +profiler_bpf__destroy(struct profiler_bpf *obj)
>>> +{
>>> +	if (!obj)
>>> +		return;
>>> +	if (obj->skeleton)
>>> +		bpf_object__destroy_skeleton(obj->skeleton);
>>> +	free(obj);
>>> +}
>>> +
>>> +static inline int
>>> +profiler_bpf__create_skeleton(struct profiler_bpf *obj);
>>> +
>>> +static inline struct profiler_bpf *
>>> +profiler_bpf__open_opts(const struct bpf_object_open_opts *opts)
>>> +{
>>> +	struct profiler_bpf *obj;
>>> +
>>> +	obj = (typeof(obj))calloc(1, sizeof(*obj));
>>> +	if (!obj)
>>> +		return NULL;
>>> +	if (profiler_bpf__create_skeleton(obj))
>>> +		goto err;
>>> +	if (bpf_object__open_skeleton(obj->skeleton, opts))
>>> +		goto err;
>>> +
>>> +	return obj;
>>> +err:
>>> +	profiler_bpf__destroy(obj);
>>> +	return NULL;
>>> +}
>>> +
>>> +static inline struct profiler_bpf *
>>> +profiler_bpf__open(void)
>>> +{
>>> +	return profiler_bpf__open_opts(NULL);
>>> +}
>>> +
>>> +static inline int
>>> +profiler_bpf__load(struct profiler_bpf *obj)
>>> +{
>>> +	return bpf_object__load_skeleton(obj->skeleton);
>>> +}
>>> +
>>> +static inline struct profiler_bpf *
>>> +profiler_bpf__open_and_load(void)
>>> +{
>>> +	struct profiler_bpf *obj;
>>> +
>>> +	obj = profiler_bpf__open();
>>> +	if (!obj)
>>> +		return NULL;
>>> +	if (profiler_bpf__load(obj)) {
>>> +		profiler_bpf__destroy(obj);
>>> +		return NULL;
>>> +	}
>>> +	return obj;
>>> +}
>>> +
>>> +static inline int
>>> +profiler_bpf__attach(struct profiler_bpf *obj)
>>> +{
>>> +	return bpf_object__attach_skeleton(obj->skeleton);
>>> +}
>>> +
>>> +static inline void
>>> +profiler_bpf__detach(struct profiler_bpf *obj)
>>> +{
>>> +	return bpf_object__detach_skeleton(obj->skeleton);
>>> +}
>>> +
>>> +static inline int
>>> +profiler_bpf__create_skeleton(struct profiler_bpf *obj)
>>> +{
>>> +	struct bpf_object_skeleton *s;
>>> +
>>> +	s = (typeof(s))calloc(1, sizeof(*s));
>>> +	if (!s)
>>> +		return -1;
>>> +	obj->skeleton = s;
>>> +
>>> +	s->sz = sizeof(*s);
>>> +	s->name = "profiler_bpf";
>>> +	s->obj = &obj->obj;
>>> +
>>> +	/* maps */
>>> +	s->map_cnt = 6;
>>> +	s->map_skel_sz = sizeof(*s->maps);
>>> +	s->maps = (typeof(s->maps))calloc(s->map_cnt, s->map_skel_sz);
>>> +	if (!s->maps)
>>> +		goto err;
>>> +
>>> +	s->maps[0].name = "events";
>>> +	s->maps[0].map = &obj->maps.events;
>>> +
>>> +	s->maps[1].name = "fentry_readings";
>>> +	s->maps[1].map = &obj->maps.fentry_readings;
>>> +
>>> +	s->maps[2].name = "accum_readings";
>>> +	s->maps[2].map = &obj->maps.accum_readings;
>>> +
>>> +	s->maps[3].name = "counts";
>>> +	s->maps[3].map = &obj->maps.counts;
>>> +
>>> +	s->maps[4].name = "miss_counts";
>>> +	s->maps[4].map = &obj->maps.miss_counts;
>>> +
>>> +	s->maps[5].name = "profiler.rodata";
>>> +	s->maps[5].map = &obj->maps.rodata;
>>> +	s->maps[5].mmaped = (void **)&obj->rodata;
>>> +
>>> +	/* programs */
>>> +	s->prog_cnt = 2;
>>> +	s->prog_skel_sz = sizeof(*s->progs);
>>> +	s->progs = (typeof(s->progs))calloc(s->prog_cnt, s->prog_skel_sz);
>>> +	if (!s->progs)
>>> +		goto err;
>>> +
>>> +	s->progs[0].name = "fentry_XXX";
>>> +	s->progs[0].prog = &obj->progs.fentry_XXX;
>>> +	s->progs[0].link = &obj->links.fentry_XXX;
>>> +
>>> +	s->progs[1].name = "fexit_XXX";
>>> +	s->progs[1].prog = &obj->progs.fexit_XXX;
>>> +	s->progs[1].link = &obj->links.fexit_XXX;
>>> +
>>> +	s->data_sz = 18256;
>>> +	s->data = (void *)"\
>>> +\x7f\x45\x4c\x46\x02\x01\x01\0\0\0\0\0\0\0\0\0\x01\0\xf7\0\x01\0\0\0\0\0\0\0\0\
>> 
>> Holy binary blob, Batman! :)
>> 
>> What is this blob, exactly? The bytecode output of a precompiled
>> program?
>
> It is the skeleton compiled from profiler.bpf.c. Please refer to 
> the README file for step to generate it.

Ah, right, seems I managed to skip over the second half of the patch so
missed that :)

> In long term, we should include the generation of this blob in the 
> make process. But for RFC, I kept it simple for now. :)

Yes, I agree; but fair enough.

-Toke

Powered by blists - more mailing lists