Re: [PATCH v4 perf,bpf 06/15] perf, bpf: save bpf_prog_info in a rbtree in perf_env

From: Song Liu <songliubraving@fb.com>
To: Jiri Olsa <jolsa@redhat.com>
Cc: Netdev <netdev@vger.kernel.org>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	"ast@kernel.org" <ast@kernel.org>,
	"daniel@iogearbox.net" <daniel@iogearbox.net>,
	Kernel Team <Kernel-team@fb.com>,
	"peterz@infradead.org" <peterz@infradead.org>,
	"acme@redhat.com" <acme@redhat.com>,
	"jolsa@kernel.org" <jolsa@kernel.org>,
	"namhyung@kernel.org" <namhyung@kernel.org>
Subject: Re: [PATCH v4 perf,bpf 06/15] perf, bpf: save bpf_prog_info in a rbtree in perf_env
Date: Wed, 27 Feb 2019 17:38:03 +0000	[thread overview]
Message-ID: <B857ABD1-046A-4B1F-B498-C09273B1FD5E@fb.com> (raw)
In-Reply-To: <20190227132123.GE18893@krava>

> On Feb 27, 2019, at 5:21 AM, Jiri Olsa <jolsa@redhat.com> wrote:
> 
> On Mon, Feb 25, 2019 at 04:20:10PM -0800, Song Liu wrote:
>> bpf_prog_info contains information necessary to annotate bpf programs.
>> This patch saves bpf_prog_info for bpf programs loaded in the system.
>> 
>> Signed-off-by: Song Liu <songliubraving@fb.com>
>> ---
>> tools/perf/util/bpf-event.c | 32 +++++++++++++-
>> tools/perf/util/bpf-event.h |  7 ++-
>> tools/perf/util/env.c       | 85 +++++++++++++++++++++++++++++++++++++
>> tools/perf/util/env.h       | 17 ++++++++
>> 4 files changed, 139 insertions(+), 2 deletions(-)
>> 
>> diff --git a/tools/perf/util/bpf-event.c b/tools/perf/util/bpf-event.c
>> index ff7ee149ec46..ce81b2c43a51 100644
>> --- a/tools/perf/util/bpf-event.c
>> +++ b/tools/perf/util/bpf-event.c
>> @@ -10,6 +10,7 @@
>> #include "debug.h"
>> #include "symbol.h"
>> #include "machine.h"
>> +#include "env.h"
>> #include "session.h"
>> 
>> #define ptr_to_u64(ptr)    ((__u64)(unsigned long)(ptr))
>> @@ -54,17 +55,28 @@ static int perf_event__synthesize_one_bpf_prog(struct perf_session *session,
>> 	struct bpf_event *bpf_event = &event->bpf_event;
>> 	struct bpf_prog_info_linear *info_linear;
>> 	struct perf_tool *tool = session->tool;
>> +	struct bpf_prog_info_node *info_node;
>> 	struct bpf_prog_info *info;
>> 	struct btf *btf = NULL;
>> 	bool has_btf = false;
>> +	struct perf_env *env;
>> 	u32 sub_prog_cnt, i;
>> 	int err = 0;
>> 	u64 arrays;
>> 
>> +	/*
>> +	 * for perf-record and perf-report use header.env;
>> +	 * otherwise, use global perf_env.
>> +	 */
>> +	env = session->data ? &session->header.env : &perf_env;
>> +
>> 	arrays = 1UL << BPF_PROG_INFO_JITED_KSYMS;
>> 	arrays |= 1UL << BPF_PROG_INFO_JITED_FUNC_LENS;
>> 	arrays |= 1UL << BPF_PROG_INFO_FUNC_INFO;
>> 	arrays |= 1UL << BPF_PROG_INFO_PROG_TAGS;
>> +	arrays |= 1UL << BPF_PROG_INFO_JITED_INSNS;
>> +	arrays |= 1UL << BPF_PROG_INFO_LINE_INFO;
>> +	arrays |= 1UL << BPF_PROG_INFO_JITED_LINE_INFO;
>> 
>> 	info_linear = bpf_program__get_prog_info_linear(fd, arrays);
>> 	if (IS_ERR_OR_NULL(info_linear)) {
>> @@ -153,8 +165,8 @@ static int perf_event__synthesize_one_bpf_prog(struct perf_session *session,
>> 						     machine, process);
>> 	}
>> 
>> -	/* Synthesize PERF_RECORD_BPF_EVENT */
>> 	if (opts->bpf_event) {
>> +		/* Synthesize PERF_RECORD_BPF_EVENT */
>> 		*bpf_event = (struct bpf_event){
>> 			.header = {
>> 				.type = PERF_RECORD_BPF_EVENT,
>> @@ -167,6 +179,24 @@ static int perf_event__synthesize_one_bpf_prog(struct perf_session *session,
>> 		memcpy(bpf_event->tag, info->tag, BPF_TAG_SIZE);
>> 		memset((void *)event + event->header.size, 0, machine->id_hdr_size);
>> 		event->header.size += machine->id_hdr_size;
>> +
>> +		/* save bpf_prog_info to env */
> 
> why do we save this to perf_env in here? we just
> synthesize the same data as event, so report and
> top will read it and fill perf_env again, right?

The synthesized events are same as PERF_RECORD_BPF_EVENT
PROG_LOAD. We need to process them and fill perf_env as
soon as we get PROG_LOAD. Otherwise, the program might 
unload before we process them. 

> 
> could you please explain the whole flow of the
> bpf events and its respective data in perf_env
> and put it into the changelog

I will add more information to the change log. 

> 
>> +		info_node = malloc(sizeof(struct bpf_prog_info_node));
>> +
>> +		/*
>> +		 * Do not bail out for !info_node, as we still want to
>> +		 * call  perf_tool__process_synth_event()
> 
> well, we are out of memory, so I dont think perf_tool__process_synth_event
> will get too far.. also the perf_env data would be inconsistent with what
> you store as event.. how can that work?

It is OK we have PERF_RECORD_BPF_EVENT but not related
bpf_prog_info and btf. The perf.data file will show 
a BPF program was loaded, but we won't be able to do 
annotation. Does this make sense?

Thanks,
Song

> 
> thanks,
> jirka
> 
>> +		 */
>> +		if (info_node) {
>> +			info_node->info_linear = info_linear;
>> +			perf_env__insert_bpf_prog_info(env, info_node);
>> +			info_linear = NULL;
>> +		}
>> +
> 
> SNIP