Re: [RFC/PATCH 09/38] perf record: Add --index option for building index table

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Fri, Oct 02, 2015 at 02:18:50PM +0900, Namhyung Kim escreveu:
> The new --index option will create indexed data file which can be
> processed by multiple threads parallelly.  It saves meta event and
> sample data in separate files and merges them with an index table.
> 
> If there's an index table in the data file, the HEADER_DATA_INDEX
> feature bit is set and session->header.index[0] will point to the meta
> event area, and rest are sample data.  It'd look like below:

So this is all about perf.data files, i.e. we will traverse it all
looking for metadata events, then the samples itself will be processed
using multiple threads. I.e. two stages, touching the whole perf.data
file looking for the metadata events, then touching it all again to
process the files, right?

The model processing events for 'perf report' and for 'perf top' will
differ, right?

Can't we have one thread reading the events, another batching then to
merge up to those FINISHED_ROUND when it needs to sort things (in
parallel with reading more events for the next round) and then pass the
sorted batch to one thread per CPU to actually process the samples as in
hist processing, etc?

Need to try this, see if 'perf top' works, which I think it will, but
you haven't mentioned anything about it in the cover letter for this
patchkit.

But no speedups should be expected there, as no 'perf.data' file is
involved...

- Arnaldo
 
>         +---------------------+
>         |     file header     |
>         |---------------------|
>         |                     |
>         |    meta events[0] <-+--+
>         |                     |  |
>         |---------------------|  |
>         |                     |  |
>         |    sample data[1] <-+--+
>         |                     |  |
>         |---------------------|  |
>         |                     |  |
>         |    sample data[2] <-|--+
>         |                     |  |
>         |---------------------|  |
>         |         ...         | ...
>         |---------------------|  |
>         |     feature data    |  |
>         |   (contains index) -+--+
>         +---------------------+
> 
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> ---
>  tools/perf/Documentation/perf-record.txt |   4 +
>  tools/perf/builtin-record.c              | 178 ++++++++++++++++++++++++++++---
>  tools/perf/perf.h                        |   1 +
>  tools/perf/util/header.c                 |   2 +
>  tools/perf/util/session.c                |   1 +
>  5 files changed, 173 insertions(+), 13 deletions(-)
> 
> diff --git a/tools/perf/Documentation/perf-record.txt b/tools/perf/Documentation/perf-record.txt
> index 2e9ce77b5e14..71a9520b10b0 100644
> --- a/tools/perf/Documentation/perf-record.txt
> +++ b/tools/perf/Documentation/perf-record.txt
> @@ -308,6 +308,10 @@ This option sets the time out limit. The default value is 500 ms.
>  Record context switch events i.e. events of type PERF_RECORD_SWITCH or
>  PERF_RECORD_SWITCH_CPU_WIDE.
>  
> +--index::
> +Build an index table for sample data.  This will speed up perf report by
> +parallel processing.
> +
>  SEE ALSO
>  --------
>  linkperf:perf-stat[1], linkperf:perf-list[1]
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 623984c81478..096634c4c5ea 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -43,6 +43,7 @@ struct record {
>  	u64			bytes_written;
>  	struct perf_data_file	file;
>  	struct auxtrace_record	*itr;
> +	int			*fds;
>  	struct perf_evlist	*evlist;
>  	struct perf_session	*session;
>  	const char		*progname;
> @@ -52,9 +53,16 @@ struct record {
>  	long			samples;
>  };
>  
> -static int record__write(struct record *rec, void *bf, size_t size)
> +static int record__write(struct record *rec, void *bf, size_t size, int idx)
>  {
> -	if (perf_data_file__write(rec->session->file, bf, size) < 0) {
> +	int fd;
> +
> +	if (rec->fds && idx >= 0)
> +		fd = rec->fds[idx];
> +	else
> +		fd = perf_data_file__fd(rec->session->file);
> +
> +	if (writen(fd, bf, size) < 0) {
>  		pr_err("failed to write perf data, error: %m\n");
>  		return -1;
>  	}
> @@ -69,7 +77,7 @@ static int process_synthesized_event(struct perf_tool *tool,
>  				     struct machine *machine __maybe_unused)
>  {
>  	struct record *rec = container_of(tool, struct record, tool);
> -	return record__write(rec, event, event->header.size);
> +	return record__write(rec, event, event->header.size, -1);
>  }
>  
>  static int record__mmap_read(struct record *rec, int idx)
> @@ -94,7 +102,7 @@ static int record__mmap_read(struct record *rec, int idx)
>  		size = md->mask + 1 - (old & md->mask);
>  		old += size;
>  
> -		if (record__write(rec, buf, size) < 0) {
> +		if (record__write(rec, buf, size, idx) < 0) {
>  			rc = -1;
>  			goto out;
>  		}
> @@ -104,7 +112,7 @@ static int record__mmap_read(struct record *rec, int idx)
>  	size = head - old;
>  	old += size;
>  
> -	if (record__write(rec, buf, size) < 0) {
> +	if (record__write(rec, buf, size, idx) < 0) {
>  		rc = -1;
>  		goto out;
>  	}
> @@ -151,6 +159,7 @@ static int record__process_auxtrace(struct perf_tool *tool,
>  	struct perf_data_file *file = &rec->file;
>  	size_t padding;
>  	u8 pad[8] = {0};
> +	int idx = event->auxtrace.idx;
>  
>  	if (!perf_data_file__is_pipe(file)) {
>  		off_t file_offset;
> @@ -171,11 +180,11 @@ static int record__process_auxtrace(struct perf_tool *tool,
>  	if (padding)
>  		padding = 8 - padding;
>  
> -	record__write(rec, event, event->header.size);
> -	record__write(rec, data1, len1);
> +	record__write(rec, event, event->header.size, idx);
> +	record__write(rec, data1, len1, idx);
>  	if (len2)
> -		record__write(rec, data2, len2);
> -	record__write(rec, &pad, padding);
> +		record__write(rec, data2, len2, idx);
> +	record__write(rec, &pad, padding, idx);
>  
>  	return 0;
>  }
> @@ -268,6 +277,110 @@ int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
>  
>  #endif
>  
> +#define INDEX_FILE_FMT  "%s.dir/perf.data.%d"
> +
> +static int record__create_index_files(struct record *rec, int nr_index)
> +{
> +	int i = 0;
> +	int ret = -1;
> +	char path[PATH_MAX];
> +	struct perf_data_file *file = &rec->file;
> +
> +	rec->fds = malloc(nr_index * sizeof(int));
> +	if (rec->fds == NULL)
> +		return -ENOMEM;
> +
> +	scnprintf(path, sizeof(path), "%s.dir", file->path);
> +	if (rm_rf(path) < 0 || mkdir(path, S_IRWXU) < 0)
> +		goto out_err;
> +
> +	for (i = 0; i < nr_index; i++) {
> +		scnprintf(path, sizeof(path), INDEX_FILE_FMT, file->path, i);
> +		ret = open(path, O_RDWR|O_CREAT|O_TRUNC, S_IRUSR|S_IWUSR);
> +		if (ret < 0)
> +			goto out_err;
> +
> +		rec->fds[i] = ret;
> +	}
> +	return 0;
> +
> +out_err:
> +	while (--i >= 1)
> +		close(rec->fds[i]);
> +	zfree(&rec->fds);
> +
> +	scnprintf(path, sizeof(path), "%s.dir", file->path);
> +	rm_rf(path);
> +
> +	return ret;
> +}
> +
> +static int record__merge_index_files(struct record *rec, int nr_index)
> +{
> +	int i;
> +	int ret = -ENOMEM;
> +	u64 offset;
> +	char path[PATH_MAX];
> +	struct perf_file_section *idx;
> +	struct perf_data_file *file = &rec->file;
> +	struct perf_session *session = rec->session;
> +	int output_fd = perf_data_file__fd(file);
> +
> +	/* +1 for header file itself */
> +	nr_index++;
> +
> +	idx = calloc(nr_index, sizeof(*idx));
> +	if (idx == NULL)
> +		goto out_close;
> +
> +	offset = lseek(output_fd, 0, SEEK_END);
> +
> +	idx[0].offset = session->header.data_offset;
> +	idx[0].size   = offset - idx[0].offset;
> +
> +	for (i = 1; i < nr_index; i++) {
> +		struct stat stbuf;
> +		int fd = rec->fds[i - 1];
> +
> +		ret = fstat(fd, &stbuf);
> +		if (ret < 0)
> +			goto out_close;
> +
> +		idx[i].offset = offset;
> +		idx[i].size   = stbuf.st_size;
> +
> +		offset += stbuf.st_size;
> +
> +		if (idx[i].size == 0)
> +			continue;
> +
> +		ret = copyfile_offset(fd, 0, output_fd, idx[i].offset,
> +				      idx[i].size);
> +		if (ret < 0)
> +			goto out_close;
> +	}
> +
> +	session->header.index = idx;
> +	session->header.nr_index = nr_index;
> +
> +	perf_has_index = true;
> +
> +	ret = 0;
> +
> +out_close:
> +	if (ret < 0)
> +		pr_err("failed to merge index files: %d\n", ret);
> +
> +	for (i = 0; i < nr_index - 1; i++)
> +		close(rec->fds[i]);
> +
> +	scnprintf(path, sizeof(path), "%s.dir", file->path);
> +	rm_rf(path);
> +
> +	zfree(&rec->fds);
> +	return ret;
> +}
> +
>  static int record__open(struct record *rec)
>  {
>  	char msg[512];
> @@ -306,7 +419,8 @@ try_again:
>  
>  	if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
>  				 opts->auxtrace_mmap_pages,
> -				 opts->auxtrace_snapshot_mode, false) < 0) {
> +				 opts->auxtrace_snapshot_mode,
> +				 opts->index) < 0) {
>  		if (errno == EPERM) {
>  			pr_err("Permission error mapping pages.\n"
>  			       "Consider increasing "
> @@ -323,6 +437,14 @@ try_again:
>  		goto out;
>  	}
>  
> +	if (opts->index) {
> +		rc = record__create_index_files(rec, evlist->nr_mmaps);
> +		if (rc < 0) {
> +			pr_err("failed to create index file: %d\n", rc);
> +			goto out;
> +		}
> +	}
> +
>  	session->evlist = evlist;
>  	perf_session__set_id_hdr_size(session);
>  out:
> @@ -347,7 +469,9 @@ static int process_buildids(struct record *rec)
>  	struct perf_data_file *file  = &rec->file;
>  	struct perf_session *session = rec->session;
>  
> -	if (file->size == 0)
> +	/* update file size after merging sample files with index */
> +	u64 size = lseek(perf_data_file__fd(file), 0, SEEK_END);
> +	if (size == 0)
>  		return 0;
>  
>  	/*
> @@ -414,6 +538,13 @@ static int record__mmap_read_all(struct record *rec)
>  			}
>  		}
>  
> +		if (rec->evlist->track_mmap && rec->evlist->track_mmap[i].base) {
> +			if (record__mmap_read(rec, track_mmap_idx(i)) != 0) {
> +				rc = -1;
> +				goto out;
> +			}
> +		}
> +
>  		if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
>  		    record__auxtrace_mmap_read(rec, mm) != 0) {
>  			rc = -1;
> @@ -426,7 +557,8 @@ static int record__mmap_read_all(struct record *rec)
>  	 * at least one event.
>  	 */
>  	if (bytes_written != rec->bytes_written)
> -		rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
> +		rc = record__write(rec, &finished_round_event,
> +				   sizeof(finished_round_event), -1);
>  
>  out:
>  	return rc;
> @@ -452,7 +584,8 @@ static void record__init_features(struct record *rec)
>  	if (!rec->opts.full_auxtrace)
>  		perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
>  
> -	perf_header__clear_feat(&session->header, HEADER_DATA_INDEX);
> +	if (!rec->opts.index)
> +		perf_header__clear_feat(&session->header, HEADER_DATA_INDEX);
>  }
>  
>  static volatile int workload_exec_errno;
> @@ -520,6 +653,11 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
>  		}
>  	}
>  
> +	if (file->is_pipe && opts->index) {
> +		pr_warning("Indexing is disabled for pipe output\n");
> +		opts->index = false;
> +	}
> +
>  	if (record__open(rec) != 0) {
>  		err = -1;
>  		goto out_child;
> @@ -753,6 +891,9 @@ out_child:
>  		rec->session->header.data_size += rec->bytes_written;
>  		file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
>  
> +		if (rec->opts.index)
> +			record__merge_index_files(rec, rec->evlist->nr_mmaps);
> +
>  		if (!rec->no_buildid) {
>  			process_buildids(rec);
>  			/*
> @@ -1119,6 +1260,8 @@ struct option __record_options[] = {
>  			"per thread proc mmap processing timeout in ms"),
>  	OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
>  		    "Record context switch events"),
> +	OPT_BOOLEAN(0, "index", &record.opts.index,
> +		    "make index for sample data to speed-up processing"),
>  	OPT_END()
>  };
>  
> @@ -1186,6 +1329,15 @@ int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
>  		goto out_symbol_exit;
>  	}
>  
> +	if (rec->opts.index) {
> +		if (!rec->opts.sample_time) {
> +			pr_err("Sample timestamp is required for indexing\n");
> +			goto out_symbol_exit;
> +		}
> +
> +		perf_evlist__add_dummy_tracking(rec->evlist);
> +	}
> +
>  	if (rec->opts.target.tid && !rec->opts.no_inherit_set)
>  		rec->opts.no_inherit = true;
>  
> diff --git a/tools/perf/perf.h b/tools/perf/perf.h
> index f4b4d7d8752c..df7c208abb74 100644
> --- a/tools/perf/perf.h
> +++ b/tools/perf/perf.h
> @@ -60,6 +60,7 @@ struct record_opts {
>  	bool	     full_auxtrace;
>  	bool	     auxtrace_snapshot_mode;
>  	bool	     record_switch_events;
> +	bool	     index;
>  	unsigned int freq;
>  	unsigned int mmap_pages;
>  	unsigned int auxtrace_mmap_pages;
> diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c
> index c357f7f47d32..13ba1402ec1b 100644
> --- a/tools/perf/util/header.c
> +++ b/tools/perf/util/header.c
> @@ -2706,6 +2706,8 @@ int perf_session__read_header(struct perf_session *session)
>  						   session->tevent.pevent))
>  		goto out_delete_evlist;
>  
> +	perf_has_index = perf_header__has_feat(&session->header, HEADER_DATA_INDEX);
> +
>  	return 0;
>  out_errno:
>  	return -errno;
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index 91fa9647f565..7546c4d147b9 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -182,6 +182,7 @@ void perf_session__delete(struct perf_session *session)
>  	machines__exit(&session->machines);
>  	if (session->file)
>  		perf_data_file__close(session->file);
> +	free(session->header.index);
>  	free(session);
>  }
>  
> -- 
> 2.6.0