Re: [PATCH v2 perf,bpf 11/11] perf, bpf: save information about short living bpf programs

[Song Liu]

> On Feb 15, 2019, at 6:41 AM, Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> 
> Em Thu, Feb 14, 2019 at 04:00:45PM -0800, Song Liu escreveu:
>> To annotate bpf programs in perf, it is necessary to save information in
>> bpf_prog_info and btf. For short living bpf program, it is necessary to
>> save these information before it is unloaded.
>> 
>> This patch saves these information in a separate thread. This thread
>> creates its own evlist, that only tracks bpf events. This evlists uses
>> ring buffer with very low watermark for lower latency. When bpf load
>> events are received, this thread tries to gather information via sys_bpf
>> and save it in perf_env.
>> 
>> Signed-off-by: Song Liu <songliubraving@xxxxxx>
>> ---
>> tools/perf/builtin-record.c |  13 ++++
>> tools/perf/builtin-top.c    |  12 ++++
>> tools/perf/util/bpf-event.c | 129 ++++++++++++++++++++++++++++++++++++
>> tools/perf/util/bpf-event.h |  22 ++++++
>> tools/perf/util/evlist.c    |  20 ++++++
>> tools/perf/util/evlist.h    |   2 +
>> 6 files changed, 198 insertions(+)
>> 
>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>> index 2355e0a9eda0..46abb44aaaab 100644
>> --- a/tools/perf/builtin-record.c
>> +++ b/tools/perf/builtin-record.c
>> @@ -1106,6 +1106,8 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
>> 	struct perf_data *data = &rec->data;
>> 	struct perf_session *session;
>> 	bool disabled = false, draining = false;
>> +	struct bpf_event_poll_args poll_args;
>> +	bool bpf_thread_running = false;
>> 	int fd;
>> 
>> 	atexit(record__sig_exit);
>> @@ -1206,6 +1208,14 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
>> 		goto out_child;
>> 	}
>> 
>> +	if (rec->opts.bpf_event) {
>> +		poll_args.env = &session->header.env;
>> +		poll_args.target = &rec->opts.target;
>> +		poll_args.done = &done;
>> +		if (bpf_event__start_polling_thread(&poll_args) == 0)
>> +			bpf_thread_running = true;
>> +	}
>> +
>> 	err = record__synthesize(rec, false);
>> 	if (err < 0)
>> 		goto out_child;
>> @@ -1456,6 +1466,9 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
>> 
>> out_delete_session:
>> 	perf_session__delete(session);
>> +
>> +	if (bpf_thread_running)
>> +		bpf_event__stop_polling_thread(&poll_args);
>> 	return status;
>> }
>> 
>> diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
>> index 5271d7211b9c..2586ee081967 100644
>> --- a/tools/perf/builtin-top.c
>> +++ b/tools/perf/builtin-top.c
>> @@ -1524,10 +1524,12 @@ int cmd_top(int argc, const char **argv)
>> 			"number of thread to run event synthesize"),
>> 	OPT_END()
>> 	};
>> +	struct bpf_event_poll_args poll_args;
>> 	const char * const top_usage[] = {
>> 		"perf top [<options>]",
>> 		NULL
>> 	};
>> +	bool bpf_thread_running = false;
>> 	int status = hists__init();
>> 
>> 	if (status < 0)
>> @@ -1652,8 +1654,18 @@ int cmd_top(int argc, const char **argv)
>> 		signal(SIGWINCH, winch_sig);
>> 	}
>> 
>> +	if (top.record_opts.bpf_event) {
>> +		poll_args.env = &perf_env;
>> +		poll_args.target = target;
>> +		poll_args.done = &done;
>> +		if (bpf_event__start_polling_thread(&poll_args) == 0)
>> +			bpf_thread_running = true;
>> +	}
>> 	status = __cmd_top(&top);
>> 
>> +	if (bpf_thread_running)
>> +		bpf_event__stop_polling_thread(&poll_args);
>> +
>> out_delete_evlist:
>> 	perf_evlist__delete(top.evlist);
>> 
>> diff --git a/tools/perf/util/bpf-event.c b/tools/perf/util/bpf-event.c
>> index 4f347d61ed96..0caf137c515b 100644
>> --- a/tools/perf/util/bpf-event.c
>> +++ b/tools/perf/util/bpf-event.c
>> @@ -8,6 +8,7 @@
>> #include "machine.h"
>> #include "env.h"
>> #include "session.h"
>> +#include "evlist.h"
>> 
>> #define ptr_to_u64(ptr)    ((__u64)(unsigned long)(ptr))
>> 
>> @@ -316,3 +317,131 @@ int perf_event__synthesize_bpf_events(struct perf_session *session,
>> 	free(event);
>> 	return err;
>> }
>> +
>> +static void perf_env_add_bpf_info(struct perf_env *env, u32 id)
>> +{
>> +	struct bpf_prog_info_linear *info_linear;
>> +	struct bpf_prog_info_node *info_node;
>> +	struct btf *btf = NULL;
>> +	u64 arrays;
>> +	u32 btf_id;
>> +	int fd;
>> +
>> +	fd = bpf_prog_get_fd_by_id(id);
>> +	if (fd < 0)
>> +		return;
>> +
>> +	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)) {
>> +		pr_debug("%s: failed to get BPF program info. aborting\n", __func__);
>> +		goto out;
>> +	}
>> +
>> +	btf_id = info_linear->info.btf_id;
>> +
>> +	info_node = malloc(sizeof(struct bpf_prog_info_node));
>> +	if (info_node) {
>> +		info_node->info_linear = info_linear;
>> +		perf_env__insert_bpf_prog_info(env, info_node);
>> +	} else
>> +		free(info_linear);
>> +
>> +	if (btf_id == 0)
>> +		goto out;
>> +
>> +	if (btf__get_from_id(btf_id, &btf)) {
>> +		pr_debug("%s: failed to get BTF of id %u, aborting\n",
>> +			 __func__, btf_id);
>> +		goto out;
>> +	}
>> +	perf_fetch_btf(env, btf_id, btf);
>> +
>> +out:
>> +	free(btf);
>> +	close(fd);
>> +}
>> +
>> +static void *bpf_poll_thread(void *arg)
>> +{
>> +	struct bpf_event_poll_args *args = arg;
>> +	int i;
>> +
>> +	while (!*(args->done)) {
>> +		perf_evlist__poll(args->evlist, 1000);
>> +
>> +		for (i = 0; i < args->evlist->nr_mmaps; i++) {
>> +			struct perf_mmap *map = &args->evlist->mmap[i];
>> +			union perf_event *event;
>> +
>> +			if (perf_mmap__read_init(map))
>> +				continue;
>> +			while ((event = perf_mmap__read_event(map)) != NULL) {
>> +				pr_debug("processing vip event of type %d\n",
>> +					 event->header.type);
>> +				switch (event->header.type) {
>> +				case PERF_RECORD_BPF_EVENT:
>> +					if (event->bpf_event.type != PERF_BPF_EVENT_PROG_LOAD)
>> +						break;
>> +					perf_env_add_bpf_info(args->env, event->bpf_event.id);
>> +					break;
>> +				default:
>> +					break;
>> +				}
>> +				perf_mmap__consume(map);
>> +			}
>> +			perf_mmap__read_done(map);
>> +		}
>> +	}
>> +	return NULL;
>> +}
>> +
>> +pthread_t poll_thread;
>> +
>> +int bpf_event__start_polling_thread(struct bpf_event_poll_args *args)
>> +{
>> +	struct perf_evsel *counter;
>> +
>> +	args->evlist = perf_evlist__new();
>> +
>> +	if (args->evlist == NULL)
>> +		return -1;
>> +
>> +	if (perf_evlist__create_maps(args->evlist, args->target))
> 		goto out_delete_evlist;
>> +
>> +	if (perf_evlist__add_bpf_tracker(args->evlist))
> 		goto out_delete_evlist;
>> +
>> +	evlist__for_each_entry(args->evlist, counter) {
>> +		if (perf_evsel__open(counter, args->evlist->cpus,
>> +				     args->evlist->threads) < 0)
> 			goto out_delete_evlist;
>> +	}
>> +
>> +	if (perf_evlist__mmap(args->evlist, UINT_MAX))
> 		goto out_delete_evlist;
>> +
>> +	evlist__for_each_entry(args->evlist, counter) {
>> +		if (perf_evsel__enable(counter))
> 			goto out_delete_evlist;
>> +	}
>> +
>> +	if (pthread_create(&poll_thread, NULL, bpf_poll_thread, args))
> 		goto out_delete_evlist;	
>> +
>> +	return 0;
> out_delete_evlist:
> 	perf_evlist__delete(args->evlist);
> 	args->evlist = NULL;
> 
> 	return -1;
>> +}
>> +
>> +void bpf_event__stop_polling_thread(struct bpf_event_poll_args *args)
>> +{
>> +	pthread_join(poll_thread, NULL);
>> +	perf_evlist__exit(args->evlist);
>> +}
>> diff --git a/tools/perf/util/bpf-event.h b/tools/perf/util/bpf-event.h
>> index c4f0f1395ea5..61914827c1e3 100644
>> --- a/tools/perf/util/bpf-event.h
>> +++ b/tools/perf/util/bpf-event.h
>> @@ -12,12 +12,17 @@
>> #include <bpf/libbpf.h>
>> #include <linux/btf.h>
>> #include <linux/rbtree.h>
>> +#include <pthread.h>
>> +#include <api/fd/array.h>
>> #include "event.h"
>> 
>> struct machine;
>> union perf_event;
>> +struct perf_env;
>> struct perf_sample;
>> struct record_opts;
>> +struct evlist;
>> +struct target;
>> 
>> struct bpf_prog_info_node {
>> 	struct bpf_prog_info_linear	*info_linear;
>> @@ -31,6 +36,13 @@ struct btf_node {
>> 	char		data[];
>> };
>> 
>> +struct bpf_event_poll_args {
>> +	struct perf_env		*env;
>> +	struct perf_evlist	*evlist;
>> +	struct target		*target;
>> +	volatile int		*done;
>> +};
>> +
>> #ifdef HAVE_LIBBPF_SUPPORT
>> int machine__process_bpf_event(struct machine *machine, union perf_event *event,
>> 			       struct perf_sample *sample);
>> @@ -39,6 +51,8 @@ int perf_event__synthesize_bpf_events(struct perf_session *session,
>> 				      perf_event__handler_t process,
>> 				      struct machine *machine,
>> 				      struct record_opts *opts);
>> +int bpf_event__start_polling_thread(struct bpf_event_poll_args *args);
>> +void bpf_event__stop_polling_thread(struct bpf_event_poll_args *args);
>> #else
>> static inline int machine__process_bpf_event(struct machine *machine __maybe_unused,
>> 					     union perf_event *event __maybe_unused,
>> @@ -54,5 +68,13 @@ static inline int perf_event__synthesize_bpf_events(struct perf_session *session
>> {
>> 	return 0;
>> }
>> +
>> +static inline int bpf_event__start_polling_thread(struct bpf_event_poll_args *args __maybe_unused)
>> +{
>> +	return 0;
>> +}
>> +void bpf_event__stop_polling_thread(struct bpf_event_poll_args *args __maybe_unused)
>> +{
>> +}
>> #endif // HAVE_LIBBPF_SUPPORT
>> #endif
>> diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
>> index 8c902276d4b4..612c079579ce 100644
>> --- a/tools/perf/util/evlist.c
>> +++ b/tools/perf/util/evlist.c
>> @@ -271,6 +271,26 @@ int perf_evlist__add_dummy(struct perf_evlist *evlist)
>> 	return 0;
>> }
>> 
>> +int perf_evlist__add_bpf_tracker(struct perf_evlist *evlist)
>> +{
>> +	struct perf_event_attr attr = {
>> +		.type	          = PERF_TYPE_SOFTWARE,
>> +		.config           = PERF_COUNT_SW_DUMMY,
>> +		.watermark        = 1,
>> +		.bpf_event        = 1,
>> +		.wakeup_watermark = 1,
>> +		.size	   = sizeof(attr), /* to capture ABI version */
>> +	};
>> +	struct perf_evsel *evsel = perf_evsel__new_idx(&attr,
>> +						       evlist->nr_entries);
>> +
>> +	if (evsel == NULL)
>> +		return -ENOMEM;
>> +
>> +	perf_evlist__add(evlist, evsel);
> 
> You could use:
> 
> 	struct perf_evlist *evlist = perf_evlist__new_dummy();
> 	if (evlist != NULL) {
> 		struct perf_evsel *evsel == perf_evlist__first(evlist);
> 		evsel->attr.bpf_event = evsel->attr.watermark = evsel->attr.wakeup_watermark = 1;
> 		return 0;
> 	}
> 	return -1;

This looks cleaner. Let me fix in next version. 

> 
> Because in this case all you'll have in this evlist is the bpf tracker,
> right? The add_bpf_tracker would be handy if we would want to have a
> pre-existing evlist with some other events and wanted to add a bpf
> tracker, no?

I think all we need is a side-band evlist instead of the main evlist. May
be we should call it side-band evlist, and make it more generic?

Thanks,
Song

> 
> - Arnaldo
> 
>> +	return 0;
>> +}
>> +
>> static int perf_evlist__add_attrs(struct perf_evlist *evlist,
>> 				  struct perf_event_attr *attrs, size_t nr_attrs)
>> {
>> diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
>> index 868294491194..a2d22715188e 100644
>> --- a/tools/perf/util/evlist.h
>> +++ b/tools/perf/util/evlist.h
>> @@ -84,6 +84,8 @@ int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
>> 
>> int perf_evlist__add_dummy(struct perf_evlist *evlist);
>> 
>> +int perf_evlist__add_bpf_tracker(struct perf_evlist *evlist);
>> +
>> int perf_evlist__add_newtp(struct perf_evlist *evlist,
>> 			   const char *sys, const char *name, void *handler);
>> 
>> -- 
>> 2.17.1