perf record: Add support to store data in directory
Jiri Olsa
reviewercover perf record: Add support to store data in directory
hi,
this patchset adds the --dir option to record command (and all the other record command that overload cmd_record) that allows the data to be stored in directory with multiple data files.
It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
$ perf record --dir perf bench sched messaging $ ls -l perf.data total 344 -rw-------. 1 jolsa jolsa 43864 Jan 20 22:26 data.0 -rw-------. 1 jolsa jolsa 30464 Jan 20 22:26 data.1 -rw-------. 1 jolsa jolsa 53816 Jan 20 22:26 data.2 -rw-------. 1 jolsa jolsa 30368 Jan 20 22:26 data.3 -rw-------. 1 jolsa jolsa 40088 Jan 20 22:26 data.4 -rw-------. 1 jolsa jolsa 42592 Jan 20 22:26 data.5 -rw-------. 1 jolsa jolsa 56136 Jan 20 22:26 data.6 -rw-------. 1 jolsa jolsa 25992 Jan 20 22:26 data.7 -rw-------. 1 jolsa jolsa 8832 Jan 20 22:26 header
There's a data file created for every cpu and it's storing data for those cpu maps. The report command will read it transparently, sort it and display as single file data.
It's possible to transform directory data into standard perf.data file via simple inject command:
$ perf inject -o perf.data.file -i perf.data
The old perf fails over the directory data with following message:
$ perf report incompatible file format (rerun with -v to learn more)
I'm now testing the record threads support, so I'd like to have some agreement on the directory data support before.
It's also available in here:
git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
perf/dir
thanks,
jirka
---
Jiri Olsa (14):
perf tools: Make rm_rf to remove single file
perf session: Add process callback to reader object perf data: Move size to struct perf_data_file
perf data: Add global path holder
perf data: Make check_backup work over directories perf data: Add perf_data__(create_dir|free_dir) functions perf data: Add perf_data__open_dir_data function perf data: Add directory support
perf data: Don't store auxtrace index for directory data file perf data: Add perf_data__update_dir function
perf data: Make perf_data__size to work over directory perf session: Add __perf_session__process_dir_events function perf session: Add path to reader object
perf record: Add --dir option to store data in directory
tools/perf/Documentation/perf-record.txt | 3 ++ tools/perf/builtin-annotate.c | 4 +-- tools/perf/builtin-buildid-cache.c | 4 +-- tools/perf/builtin-buildid-list.c | 8 ++--- tools/perf/builtin-c2c.c | 4 +-- tools/perf/builtin-diff.c | 12 +++---- tools/perf/builtin-evlist.c | 4 +-- tools/perf/builtin-inject.c | 10 +++--- tools/perf/builtin-kmem.c | 2 +- tools/perf/builtin-kvm.c | 8 ++--- tools/perf/builtin-lock.c | 8 ++--- tools/perf/builtin-mem.c | 8 ++--- tools/perf/builtin-record.c | 72 +++++++++++++++++++++++++++++++------ tools/perf/builtin-report.c | 6 ++-- tools/perf/builtin-sched.c | 16 ++++----- tools/perf/builtin-script.c | 12 +++---- tools/perf/builtin-stat.c | 6 ++-- tools/perf/builtin-timechart.c | 8 ++--- tools/perf/builtin-trace.c | 8 ++--- tools/perf/util/data-convert-bt.c | 2 +- tools/perf/util/data.c | 231 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/data.h | 24 +++++++++---- tools/perf/util/mmap.h | 23 ++++++------ tools/perf/util/session.c | 117 +++++++++++++++++++++++++++++++++++++++++++++++++++++++----- tools/perf/util/util.c | 10 ++++-- 25 files changed, 479 insertions(+), 131 deletions(-)
Jiri Olsa
reviewerPATCH 1/14 perf tools: Make rm_rf to remove single file
Let rm_rf remove file if it's provided by path.
--- tools/perf/util/util.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
tools/perf/util/util.c +8 -2
@@ -125,8 +125,14 @@ int rm_rf(const char *path)char namebuf[PATH_MAX ]; dir =opendir (path); -if (dir ==NULL ) -return 0 ; +if (dir ==NULL ) { +/* +* The path does not exist or is not directory, +* so there's no harm to try remove it. This way +* rm_rf will work over single file. +*/ +return unlink (path); + }while ((d =readdir (dir)) !=NULL && !ret) {struct stat statbuf;
Signature · 1 line
2.17.2
Re: [PATCH 01/14] perf tools: Make rm_rf to remove single file
Quoted 14 lines · Jiri Olsa
Let
rm_rfremove file if it's provided by path.Signed-off-by Jiri Olsa <[email protected]>--- tools/perf/util/util.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)tools/perf/util/util.c +0
-0@@ -125,8 +125,14 @@ int rm_rf(const char *path)char namebuf[PATH_MAX ];
tools/perf/util/util.c int rm_rf(const char *path) · line 133
Quoted 10 lines
tools/perf/util/util.c +6
-2dir =opendir (path); -if (dir ==NULL ) -return 0 ; +if (dir ==NULL ) { +/* +* The path does not exist or is not directory, +* so there's no harm to try remove it. This way +* rm_rf will work over single file. +*/
This can also happen due to lack of fds or memory. Not sure file still has to be deleted in these cases.
- Alexey
Quoted 6 lines
tools/perf/util/util.c +2
-0+return unlink (path); + }while ((d =readdir (dir)) !=NULL && !ret) {struct stat statbuf;
Re: [PATCH 01/14] perf tools: Make rm_rf to remove single file
Quoted 17 lines · Alexey Budankov
Quoted 15 lines · Jiri Olsa
Let
rm_rfremove file if it's provided by path.Signed-off-by Jiri Olsa <[email protected]>--- tools/perf/util/util.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)tools/perf/util/util.c +0
-0@@ -125,8 +125,14 @@ int rm_rf(const char *path)char namebuf[PATH_MAX ];
Quoted 12 lines
tools/perf/util/util.c int rm_rf(const char *path) · line 133
Quoted 9 lines
tools/perf/util/util.c +6
-2dir =opendir (path); -if (dir ==NULL ) -return 0 ; +if (dir ==NULL ) { +/* +* The path does not exist or is not directory, +* so there's no harm to try remove it. This way +* rm_rf will work over single file. +*/ This can also happen due to lack of fds or memory. Not sure file still has to be deleted in these cases.
ok, will do proper error check
thanks,
jirka
Quoted 9 lines
- Alexey
Quoted 6 lines
tools/perf/util/util.c +2
-0+return unlink (path); + }while ((d =readdir (dir)) !=NULL && !ret) {struct stat statbuf;
Jiri Olsa
reviewerPATCH 2/14 perf session: Add process callback to reader object
In reply to Jiri OlsaAdding callback function to reader object so
callers can process data in different ways.
--- tools/perf/util/session.c | 23 +++++++++++++++++++---- 1 file changed, 19 insertions(+), 4 deletions(-)
tools/perf/util/session.c +19 -4
@@ -1835,10 +1835,17 @@ fetch_mmaped_event(struct perf_session *session,#define NUM_MMAPS 128#endif +struct reader ; + +typedef s64 (*reader_cb_t )(struct perf_session *session, +union perf_event *event, +u64 file_offset); +struct reader { -int fd; -u64 data_size; -u64 data_offset; +int fd; +u64 data_size; +u64 data_offset; +reader_cb_t process; }; static int
@@ -1909,7 +1916,7 @@ reader__process_events(struct reader *rd, struct perf_session *session, size = event->header.size;if (size <sizeof (struct perf_event_header ) || - (skip =perf_session__process_event (session, event, file_pos)) <0 ) { + (skip = rd->process (session, event, file_pos)) <0 ) {pr_err ("%#" PRIx64" [%#x]: failed to process type: %d\n" , file_offset + head, event->header.size, event->header.type);
@@ -1935,12 +1942,20 @@ reader__process_events(struct reader *rd, struct perf_session *session,return err; } +static s64 process_simple (struct perf_session *session, +union perf_event *event, +u64 file_offset) +{ +return perf_session__process_event (session, event, file_offset); +} +static int __perf_session__process_events (struct perf_session *session) {struct reader rd = { .fd =perf_data__fd (session->data), .data_size = session->header.data_size, .data_offset = session->header.data_offset, + .process = process_simple, };struct ordered_events *oe = &session->ordered_events;struct perf_tool *tool = session->tool;
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 3/14 perf data: Move size to struct perf_data_file
In reply to Jiri OlsaWe are about to add support for multiple files,
so we need each file to keep its size.
--- tools/perf/builtin-record.c | 5 ++--- tools/perf/util/data.c | 2 +- tools/perf/util/data.h | 4 ++-- 3 files changed, 5 insertions(+), 6 deletions(-)
tools/perf/builtin-record.c +2 -3
@@ -659,10 +659,9 @@ static int process_sample_event(struct perf_tool *tool,static int process_buildids (struct record *rec) { -struct perf_data *data = &rec->data;struct perf_session *session = rec->session; -if (data->size ==0 ) +if (perf_data__size (&rec->data) ==0 )return 0 ;/*
@@ -850,7 +849,7 @@ record__finish_output(struct record *rec)return ; rec->session->header.data_size += rec->bytes_written; - data->size =lseek (perf_data__fd (data),0 ,SEEK_CUR ); + data->file.size =lseek (perf_data__fd (data),0 ,SEEK_CUR );if (!rec->no_buildid) {process_buildids (rec);
tools/perf/util/data.c +1 -1
@@ -82,7 +82,7 @@ static int open_file_read(struct perf_data *data) goto out_close; } - data->size = st.st_size; + data->file.size = st.st_size;return fd; out_close:
tools/perf/util/data.h +2 -2
@@ -12,13 +12,13 @@ enum perf_data_mode {struct perf_data_file {const char *path;int fd; +unsigned long size; };struct perf_data {struct perf_data_file file;bool is_pipe;bool force; -unsigned long size;enum perf_data_mode mode; };
@@ -44,7 +44,7 @@ static inline int perf_data__fd(struct perf_data *data)static inline unsigned long perf_data__size (struct perf_data *data) { -return data->size; +return data->file.size; }int perf_data__open (struct perf_data *data);
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 4/14 perf data: Add global path holder
In reply to Jiri OlsaAdding path to the struct perf_data. It will keep the configured path for the data (const char*). The path is struct perf_data_file is now dynamically allocated (duped) from it.
This scheme is useful/used in following patches where struct perf_data::path holds the directory path.
--- tools/perf/builtin-annotate.c | 4 +-- tools/perf/builtin-buildid-cache.c | 4 +-- tools/perf/builtin-buildid-list.c | 8 +++--- tools/perf/builtin-c2c.c | 4 +-- tools/perf/builtin-diff.c | 12 ++++----- tools/perf/builtin-evlist.c | 4 +-- tools/perf/builtin-inject.c | 10 +++----- tools/perf/builtin-kmem.c | 2 +- tools/perf/builtin-kvm.c | 8 +++--- tools/perf/builtin-lock.c | 8 +++--- tools/perf/builtin-mem.c | 8 +++--- tools/perf/builtin-record.c | 6 ++--- tools/perf/builtin-report.c | 6 ++--- tools/perf/builtin-sched.c | 16 +++++------- tools/perf/builtin-script.c | 12 ++++----- tools/perf/builtin-stat.c | 6 ++--- tools/perf/builtin-timechart.c | 8 +++--- tools/perf/builtin-trace.c | 8 +++--- tools/perf/util/data-convert-bt.c | 2 +- tools/perf/util/data.c | 39 +++++++++++++++++++++--------- tools/perf/util/data.h | 3 ++- 21 files changed, 86 insertions(+), 92 deletions(-)
tools/perf/builtin-annotate.c +2 -2
@@ -441,7 +441,7 @@ static int __cmd_annotate(struct perf_annotate *ann) }if (total_nr_samples ==0 ) { -ui__error ("The %s file has no samples!\n" , session->data->file.path); +ui__error ("The %s data has no samples!\n" , session->data->path); goto out; }
@@ -578,7 +578,7 @@ int cmd_annotate(int argc, const char **argv)if (quiet)perf_quiet_option (); - data.file.path = input_name; + data.path = input_name; annotate.session =perf_session__new (&data, false, &annotate.tool);if (annotate.session ==NULL )
tools/perf/builtin-buildid-cache.c +2 -2
@@ -416,8 +416,8 @@ int cmd_buildid_cache(int argc, const char **argv) nsi =nsinfo__new (ns_id);if (missing_filename) { - data.file.path = missing_filename; - data.force = force; + data.path = missing_filename; + data.force = force; session =perf_session__new (&data, false,NULL );if (session ==NULL )
tools/perf/builtin-buildid-list.c +3 -5
@@ -52,11 +52,9 @@ static int perf_session__list_build_ids(bool force, bool with_hits) {struct perf_session *session;struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = force, };symbol__elf_init ();
tools/perf/builtin-c2c.c +2 -2
@@ -2750,8 +2750,8 @@ static int perf_c2c__report(int argc, const char **argv)if (!input_name || !strlen (input_name)) input_name ="perf.data" ; - data.file.path = input_name; - data.force = symbol_conf.force; + data.path = input_name; + data.force = symbol_conf.force; err =setup_display (display);if (err)
tools/perf/builtin-diff.c +6 -6
@@ -708,7 +708,7 @@ static void data__fprintf(void)data__for_each_file (i, d)fprintf (stdout,"# [%d] %s %s\n" , - d->idx, d->data.file.path, + d->idx, d->data.path, !d->idx ?"(Baseline)" :"" );fprintf (stdout,"#\n" );
@@ -779,14 +779,14 @@ static int __cmd_diff(void)data__for_each_file (i, d) { d->session =perf_session__new (&d->data, false, &tool);if (!d->session) { -pr_err ("Failed to open %s\n" , d->data.file.path); +pr_err ("Failed to open %s\n" , d->data.path); ret =-1 ; goto out_delete; } ret =perf_session__process_events (d->session);if (ret) { -pr_err ("Failed to process %s\n" , d->data.file.path); +pr_err ("Failed to process %s\n" , d->data.path); goto out_delete; }
@@ -1289,9 +1289,9 @@ static int data_init(int argc, const char **argv)data__for_each_file (i, d) {struct perf_data *data = &d->data; - data->file.path = use_default ? defaults[i] : argv[i]; - data->mode =PERF_DATA_MODE_READ , - data->force = force, + data->path = use_default ? defaults[i] : argv[i]; + data->mode =PERF_DATA_MODE_READ , + data->force = force, d->idx = i; }
tools/perf/builtin-evlist.c +1 -3
@@ -23,9 +23,7 @@ static int __cmd_evlist(const char *file_name, struct perf_attr_details *detailsstruct perf_session *session;struct perf_evsel *pos;struct perf_data data = { - .file = { - .path = file_name, - }, + .path = file_name, .mode =PERF_DATA_MODE_READ , .force = details->force, };
tools/perf/builtin-inject.c +4 -6
@@ -770,10 +770,8 @@ int cmd_inject(int argc, const char **argv) .input_name = "-", .samples = LIST_HEAD_INIT(inject.samples), .output = { - .file = { - .path = "-", - }, - .mode = PERF_DATA_MODE_WRITE, + .path = "-", + .mode = PERF_DATA_MODE_WRITE, }, };struct perf_data data = {
@@ -786,7 +784,7 @@ int cmd_inject(int argc, const char **argv) "Inject build-ids into the output stream"),OPT_STRING ('i' ,"input" , &inject.input_name,"file" ,"input file name" ), -OPT_STRING ('o' ,"output" , &inject.output.file.path,"file" , +OPT_STRING ('o' ,"output" , &inject.output.path,"file" ,"output file name" ),OPT_BOOLEAN ('s' ,"sched-stat" , &inject.sched_stat, "Merge sched-stat and sched-switch for getting events "
@@ -834,7 +832,7 @@ int cmd_inject(int argc, const char **argv) inject.tool.ordered_events = inject.sched_stat; - data.file.path = inject.input_name; + data.path = inject.input_name; inject.session =perf_session__new (&data, true, &inject.tool);if (inject.session ==NULL )return -1 ;
tools/perf/builtin-kmem.c +1 -1
@@ -1949,7 +1949,7 @@ int cmd_kmem(int argc, const char **argv)return __cmd_record (argc, argv); } - data.file.path = input_name; + data.path = input_name; kmem_session = session =perf_session__new (&data, false, &perf_kmem);if (session ==NULL )
tools/perf/builtin-kvm.c +3 -5
@@ -1080,11 +1080,9 @@ static int read_events(struct perf_kvm_stat *kvm) .ordered_events = true, };struct perf_data file = { - .file = { - .path = kvm->file_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = kvm->force, + .path = kvm->file_name, + .mode =PERF_DATA_MODE_READ , + .force = kvm->force, }; kvm->tool = eops;
tools/perf/builtin-lock.c +3 -5
@@ -866,11 +866,9 @@ static int __cmd_report(bool display_info) .ordered_events = true, };struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = force, }; session =perf_session__new (&data, false, &eops);
tools/perf/builtin-mem.c +3 -5
@@ -239,11 +239,9 @@ static int process_sample_event(struct perf_tool *tool,static int report_raw_events (struct perf_mem *mem) {struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = mem->force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = mem->force, };int ret; struct perf_session *session = perf_session__new(&data, false,
tools/perf/builtin-record.c +3 -3
@@ -917,7 +917,7 @@ record__switch_output(struct record *rec, bool at_exit)if (!quiet)fprintf (stderr,"[ perf record: Dump %s.%s ]\n" , - data->file.path, timestamp); + data->path, timestamp);/* Output tracking events */ if (!at_exit) {
@@ -1460,7 +1460,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)fprintf (stderr,"[ perf record: Captured and wrote %.3f MB %s%s%s ]\n" ,perf_data__size (data) /1024.0 /1024.0 , - data->file.path, postfix, samples); + data->path, postfix, samples); } out_delete_session:
@@ -1846,7 +1846,7 @@ static struct option __record_options[] = {OPT_STRING ('C' ,"cpu" , &record.opts.target.cpu_list,"cpu" ,"list of cpus to monitor" ),OPT_U64 ('c' ,"count" , &record.opts.user_interval,"event period to sample" ), -OPT_STRING ('o' ,"output" , &record.data.file.path,"file" , +OPT_STRING ('o' ,"output" , &record.data.path,"file" ,"output file name" ), OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit, &record.opts.no_inherit_set,
tools/perf/builtin-report.c +3 -3
@@ -882,7 +882,7 @@ static int __cmd_report(struct report *rep) rep->nr_entries +=evsel__hists (pos)->nr_entries;if (rep->nr_entries ==0 ) { -ui__error ("The %s file has no samples!\n" , data->file.path); +ui__error ("The %s data has no samples!\n" , data->path);return 0 ; }
@@ -1190,8 +1190,8 @@ int cmd_report(int argc, const char **argv) input_name ="perf.data" ; } - data.file.path = input_name; - data.force = symbol_conf.force; + data.path = input_name; + data.force = symbol_conf.force; repeat: session =perf_session__new (&data, false, &report.tool);
tools/perf/builtin-sched.c +6 -10
@@ -1785,11 +1785,9 @@ static int perf_sched__read_events(struct perf_sched *sched) };struct perf_session *session;struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = sched->force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = sched->force, };int rc =-1 ;
@@ -2958,11 +2956,9 @@ static int perf_sched__timehist(struct perf_sched *sched) { "sched:sched_migrate_task", timehist_migrate_task_event, }, };struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = sched->force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = sched->force, };struct perf_session *session;
tools/perf/builtin-script.c +5 -7
@@ -2943,10 +2943,8 @@ int find_scripts(char **scripts_array, char **scripts_path_array)DIR *scripts_dir, *lang_dir;struct perf_session *session;struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , + .path = input_name, + .mode =PERF_DATA_MODE_READ , };char *temp;int i =0 ;
@@ -3419,8 +3417,8 @@ int cmd_script(int argc, const char **argv) argc =parse_options_subcommand (argc, argv, options, script_subcommands, script_usage,PARSE_OPT_STOP_AT_NON_OPTION ); - data.file.path = input_name; - data.force = symbol_conf.force; + data.path = input_name; + data.force = symbol_conf.force;if (argc >1 && !strncmp (argv[0 ],"rec" ,strlen ("rec" ))) { rec_script_path =get_script_path (argv[1 ],RECORD_SUFFIX );
@@ -3646,7 +3644,7 @@ int cmd_script(int argc, const char **argv) goto out_delete; } - input =open (data.file.path,O_RDONLY );/* input_name */ + input =open (data.path,O_RDONLY );/* input_name */ if (input <0 ) { err = -errno;perror ("failed to open file" );
tools/perf/builtin-stat.c +3 -3
@@ -1322,7 +1322,7 @@ static int __cmd_record(int argc, const char **argv) PARSE_OPT_STOP_AT_NON_OPTION);if (output_name) - data->file.path = output_name; + data->path = output_name;if (stat_config.run_count !=1 || forever) {pr_err ("Cannot use -r option with perf stat record.\n" );
@@ -1523,8 +1523,8 @@ static int __cmd_report(int argc, const char **argv) input_name ="perf.data" ; } - perf_stat.data.file.path = input_name; - perf_stat.data.mode =PERF_DATA_MODE_READ ; + perf_stat.data.path = input_name; + perf_stat.data.mode =PERF_DATA_MODE_READ ; session =perf_session__new (&perf_stat.data, false, &perf_stat.tool);if (session ==NULL )
tools/perf/builtin-timechart.c +3 -5
@@ -1602,11 +1602,9 @@ static int __cmd_timechart(struct timechart *tchart, const char *output_name) { "syscalls:sys_exit_select", process_exit_poll }, };struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = tchart->force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = tchart->force, }; struct perf_session *session = perf_session__new(&data, false,
tools/perf/builtin-trace.c +3 -5
@@ -3132,11 +3132,9 @@ static int trace__replay(struct trace *trace) { "probe:vfs_getname", trace__vfs_getname, }, };struct perf_data data = { - .file = { - .path = input_name, - }, - .mode =PERF_DATA_MODE_READ , - .force = trace->force, + .path = input_name, + .mode =PERF_DATA_MODE_READ , + .force = trace->force, };struct perf_session *session;struct perf_evsel *evsel;
tools/perf/util/data-convert-bt.c +1 -1
@@ -1650,7 +1650,7 @@ int bt_convert__perf2ctf(const char *input, const char *path,fprintf (stderr,"[ perf data convert: Converted '%s' into CTF data '%s' ]\n" , - data.file.path, path); + data.path, path); fprintf(stderr, "[ perf data convert: Converted and wrote %.3f MB (%" PRIu64 " samples",
tools/perf/util/data.c +27 -12
@@ -19,11 +19,11 @@ static bool check_pipe(struct perf_data *data)int fd =perf_data__is_read (data) ?STDIN_FILENO :STDOUT_FILENO ; -if (!data->file.path) { +if (!data->path) {if (!fstat (fd, &st) &&S_ISFIFO (st.st_mode)) is_pipe = true; }else { -if (!strcmp (data->file.path,"-" )) +if (!strcmp (data->path,"-" )) is_pipe = true; }
@@ -37,13 +37,13 @@ static int check_backup(struct perf_data *data) {struct stat st; -if (!stat (data->file.path, &st) && st.st_size) { +if (!stat (data->path, &st) && st.st_size) {/* TODO check errors properly */ char oldname[PATH_MAX ];snprintf (oldname,sizeof (oldname),"%s.old" , - data->file.path); + data->path);unlink (oldname); -rename (data->file.path, oldname); +rename (data->path, oldname); }return 0 ;
@@ -115,8 +115,22 @@ static int open_file(struct perf_data *data) fd =perf_data__is_read (data) ?open_file_read (data) :open_file_write (data); +if (fd <0 ) { +free (data->file.path); +return -1 ; + } + data->file.fd = fd; -return fd <0 ?-1 :0 ; +return 0 ; +} + +static int open_file_dup (struct perf_data *data) +{ + data->file.path =strdup (data->path); +if (!data->file.path) +return -ENOMEM ; + +return open_file (data); }int perf_data__open (struct perf_data *data)
@@ -124,14 +138,15 @@ int perf_data__open(struct perf_data *data)if (check_pipe (data))return 0 ; -if (!data->file.path) - data->file.path ="perf.data" ; +if (!data->path) + data->path ="perf.data" ; -return open_file (data); +return open_file_dup (data); }void perf_data__close (struct perf_data *data) { +free (data->file.path);close (data->file.fd); }
@@ -159,15 +174,15 @@ int perf_data__switch(struct perf_data *data,if (perf_data__is_read (data))return -EINVAL ; -if (asprintf (&new_filepath,"%s.%s" , data->file.path, postfix) <0 ) +if (asprintf (&new_filepath,"%s.%s" , data->path, postfix) <0 )return -ENOMEM ;/* * Only fire a warning, don't return error, continue fill * original file. */ -if (rename (data->file.path, new_filepath)) -pr_warning ("Failed to rename %s to %s\n" , data->file.path, new_filepath); +if (rename (data->path, new_filepath)) +pr_warning ("Failed to rename %s to %s\n" , data->path, new_filepath);if (!at_exit) {close (data->file.fd);
tools/perf/util/data.h +2 -1
@@ -10,12 +10,13 @@ enum perf_data_mode { };struct perf_data_file { -const char *path; +char *path;int fd;unsigned long size; };struct perf_data { +const char *path;struct perf_data_file file;bool is_pipe;bool force;
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 5/14 perf data: Make check_backup work over directories
In reply to Jiri OlsaChanging check_backup to work also over directory
path. Also moving the call earlier in the path,
so it can rename also directory data.
--- tools/perf/util/data.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)
tools/perf/util/data.c +7 -4
@@ -37,12 +37,15 @@ static int check_backup(struct perf_data *data) {struct stat st; +if (perf_data__is_read (data)) +return 0 ; +if (!stat (data->path, &st) && st.st_size) {/* TODO check errors properly */ char oldname[PATH_MAX ];snprintf (oldname,sizeof (oldname),"%s.old" , data->path); -unlink (oldname); +rm_rf (oldname);rename (data->path, oldname); }
@@ -95,9 +98,6 @@ static int open_file_write(struct perf_data *data)int fd;char sbuf[STRERR_BUFSIZE ]; -if (check_backup (data)) -return -1 ; - fd =open (data->file.path,O_CREAT |O_RDWR |O_TRUNC |O_CLOEXEC ,S_IRUSR |S_IWUSR );
@@ -141,6 +141,9 @@ int perf_data__open(struct perf_data *data)if (!data->path) data->path ="perf.data" ; +if (check_backup (data)) +return -1 ; +return open_file_dup (data); }
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 6/14 perf data: Add perf_data__(create_dir|free_dir) functions
In reply to Jiri OlsaAdding perf_data__create_dir to create nr files inside struct perf_data path directory:
int perf_data__create_dir(struct perf_data *data, int nr);
and function to free that data:
void perf_data__free_dir(struct perf_data *data);
--- tools/perf/util/data.c | 47 ++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/data.h | 8 +++++++ 2 files changed, 55 insertions(+)
tools/perf/util/data.c +47 -0
@@ -7,11 +7,58 @@#include <fcntl.h> #include <unistd.h> #include <string.h> +#include <asm/bug.h> #include "data.h" #include "util.h" #include "debug.h" +static void free_dir (struct perf_data_file *files,int nr) +{ +while (--nr >=1 ) { +close (files[nr].fd); +free (files[nr].path); + } +free (files); +} + +void perf_data__free_dir (struct perf_data *data) +{ +free_dir (data->dir.files, data->dir.nr); +} + +int perf_data__create_dir (struct perf_data *data,int nr) +{ +struct perf_data_file *files =NULL ; +int i, ret =-1 ; + + files =malloc (nr *sizeof (*files)); +if (!files) +return -ENOMEM ; + + data->dir.files = files; + data->dir.nr = nr; + +for (i =0 ; i < nr; i++) { +struct perf_data_file *file = &files[i]; + +if (asprintf (&file->path,"%s/data.%d" , data->path, i) <0 ) + goto out_err; + + ret =open (file->path,O_RDWR |O_CREAT |O_TRUNC ,S_IRUSR |S_IWUSR ); +if (ret <0 ) + goto out_err; + + file->fd = ret; + } + +return 0 ; + +out_err: +free_dir (files, i); +return ret; +} +static bool check_pipe (struct perf_data *data) {struct stat st;
tools/perf/util/data.h +8 -0
@@ -21,6 +21,11 @@ struct perf_data {bool is_pipe;bool force;enum perf_data_mode mode; + +struct { +struct perf_data_file *files; +int nr; + } dir; };static inline bool perf_data__is_read (struct perf_data *data)
@@ -64,4 +69,7 @@ ssize_t perf_data_file__write(struct perf_data_file *file,int perf_data__switch (struct perf_data *data,const char *postfix,size_t pos,bool at_exit); + +int perf_data__create_dir (struct perf_data *data,int nr); +void perf_data__free_dir (struct perf_data *data);#endif /* __PERF_DATA_H */
Signature · 1 line
2.17.2
Re: [PATCH 06/14] perf data: Add perf_data__(create_dir|free_dir) functions
Quoted 25 lines · Jiri Olsa
Adding
perf_data__create_dirto create nr files inside structperf_datapath directory:
intperf_data__create_dir(struct perf_data *data, int nr);and function to free that data:
voidperf_data__free_dir(struct perf_data *data);Signed-off-by Jiri Olsa <[email protected]>--- tools/perf/util/data.c | 47 ++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/data.h | 8 +++++++ 2 files changed, 55 insertions(+)tools/perf/util/data.c +1
-0@@ -7,11 +7,58 @@#include <fcntl.h> #include <unistd.h> #include <string.h> +#include <asm/bug.h> #include "data.h"
tools/perf/util/data.c line 22
Quoted 10 lines
tools/perf/util/data.c +7
-0#include "util.h" #include "debug.h" +static void free_dir (struct perf_data_file *files,int nr) +{ +while (--nr >=1 ) { +close (files[nr].fd); +free (files[nr].path); + } +free (files);
It implements closing of created files and frees corresponding memory. However it doesn't delete the files what looks like the proper rollback from perf_data__create_dir() in case of some open() failure.
Quoted 26 lines
tools/perf/util/data.c +26
-0+} + +void perf_data__free_dir (struct perf_data *data) +{ +free_dir (data->dir.files, data->dir.nr); +} + +int perf_data__create_dir (struct perf_data *data,int nr) +{ +struct perf_data_file *files =NULL ; +int i, ret =-1 ; + + files =malloc (nr *sizeof (*files)); +if (!files) +return -ENOMEM ; + + data->dir.files = files; + data->dir.nr = nr; + +for (i =0 ; i < nr; i++) { +struct perf_data_file *file = &files[i]; + +if (asprintf (&file->path,"%s/data.%d" , data->path, i) <0 ) + goto out_err; + + ret =open (file->path,O_RDWR |O_CREAT |O_TRUNC ,S_IRUSR |S_IWUSR );
tools/perf/util/data.c line 58
Quoted 10 lines
tools/perf/util/data.c +10
-0+if (ret <0 ) + goto out_err; + + file->fd = ret; + } + +return 0 ; + +out_err: +free_dir (files, i);
It looks like this is more unlink(dir) than close_files_in_dir().
Alexey
Quoted 31 lines
tools/perf/util/data.c +3
-0+return ret; +} +static bool check_pipe (struct perf_data *data) {struct stat st;tools/perf/util/data.h +8
-0@@ -21,6 +21,11 @@ struct perf_data {bool is_pipe;bool force;enum perf_data_mode mode; + +struct { +struct perf_data_file *files; +int nr; + } dir; };static inline bool perf_data__is_read (struct perf_data *data)@@ -64,4 +69,7 @@ ssize_t perf_data_file__write(struct perf_data_file *file,int perf_data__switch (struct perf_data *data,const char *postfix,size_t pos,bool at_exit); + +int perf_data__create_dir (struct perf_data *data,int nr); +void perf_data__free_dir (struct perf_data *data);#endif /* __PERF_DATA_H */
Re: [PATCH 06/14] perf data: Add perf_data__(create_dir|free_dir) functions
Quoted 30 lines · Alexey Budankov
Quoted 28 lines · Jiri Olsa
Adding
perf_data__create_dirto create nr files inside structperf_datapath directory:
intperf_data__create_dir(struct perf_data *data, int nr);and function to free that data:
voidperf_data__free_dir(struct perf_data *data);Signed-off-by Jiri Olsa <[email protected]>--- tools/perf/util/data.c | 47 ++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/data.h | 8 +++++++ 2 files changed, 55 insertions(+)tools/perf/util/data.c +1
-0@@ -7,11 +7,58 @@#include <fcntl.h> #include <unistd.h> #include <string.h> +#include <asm/bug.h> #include "data.h" #include "util.h" #include "debug.h"
Quoted 11 lines
tools/perf/util/data.c line 22
Quoted 7 lines
tools/perf/util/data.c +7
-0+static void free_dir (struct perf_data_file *files,int nr) +{ +while (--nr >=1 ) { +close (files[nr].fd); +free (files[nr].path); + } +free (files);It implements closing of created files and frees corresponding memory. However it doesn't delete the files what looks like the proper rollback from
perf_data__create_dir()in case of someopen()failure.
we don't di rollback event for the single file:
[jolsa@krava perf]$ ./perf record -b ls Error:
You may not have permission to collect stats.
Consider tweaking /proc/sys/kernel/perf_event_paranoid, which controls use of the performance events system by unprivileged users (without CAP_SYS_ADMIN).
The current value is 2:
-1: Allow use of (almost) all events by all users
Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
>= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN
Disallow raw tracepoint access by users without CAP_SYS_ADMIN
>= 1: Disallow CPU event access by users without CAP_SYS_ADMIN >= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN
To make this setting permanent, edit /etc/sysctl.conf too, e.g.:
kernel.perf_event_paranoid = -1
[jolsa@krava perf]$ ll perf.data -rw-------. 1 jolsa jolsa 0 Feb 5 14:40 perf.data
there're multiple points where we could fail during the processing and fail record command with unfinished perf.data, but yes, I think we should address that and have some cleanup, I'll check
thanks,
jirka
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 06/14] perf data: Add perf_data__(create_dir|free_dir) functions
In reply to Jiri OlsaEm Sun, Feb 03, 2019 at 04:30:10PM +0100, Jiri Olsa escreveu:
Quoted 6 lines
Adding
perf_data__create_dirto create nr files inside structperf_datapath directory:
intperf_data__create_dir(struct perf_data *data, int nr);and function to free that data:
voidperf_data__free_dir(struct perf_data *data);
Can you please rename this to perf_data__remove_dir(), I think it is more appropriate, just like that rm_rf() thing.
- Arnaldo
Quoted 96 lines
Signed-off-by Jiri Olsa <[email protected]>--- tools/perf/util/data.c | 47 ++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/data.h | 8 +++++++ 2 files changed, 55 insertions(+)tools/perf/util/data.c +47
-0@@ -7,11 +7,58 @@#include <fcntl.h> #include <unistd.h> #include <string.h> +#include <asm/bug.h> #include "data.h" #include "util.h" #include "debug.h" +static void free_dir (struct perf_data_file *files,int nr) +{ +while (--nr >=1 ) { +close (files[nr].fd); +free (files[nr].path); + } +free (files); +} + +void perf_data__free_dir (struct perf_data *data) +{ +free_dir (data->dir.files, data->dir.nr); +} + +int perf_data__create_dir (struct perf_data *data,int nr) +{ +struct perf_data_file *files =NULL ; +int i, ret =-1 ; + + files =malloc (nr *sizeof (*files)); +if (!files) +return -ENOMEM ; + + data->dir.files = files; + data->dir.nr = nr; + +for (i =0 ; i < nr; i++) { +struct perf_data_file *file = &files[i]; + +if (asprintf (&file->path,"%s/data.%d" , data->path, i) <0 ) + goto out_err; + + ret =open (file->path,O_RDWR |O_CREAT |O_TRUNC ,S_IRUSR |S_IWUSR ); +if (ret <0 ) + goto out_err; + + file->fd = ret; + } + +return 0 ; + +out_err: +free_dir (files, i); +return ret; +} +static bool check_pipe (struct perf_data *data) {struct stat st;tools/perf/util/data.h +8
-0@@ -21,6 +21,11 @@ struct perf_data {bool is_pipe;bool force;enum perf_data_mode mode; + +struct { +struct perf_data_file *files; +int nr; + } dir; };static inline bool perf_data__is_read (struct perf_data *data)@@ -64,4 +69,7 @@ ssize_t perf_data_file__write(struct perf_data_file *file,int perf_data__switch (struct perf_data *data,const char *postfix,size_t pos,bool at_exit); + +int perf_data__create_dir (struct perf_data *data,int nr); +void perf_data__free_dir (struct perf_data *data);#endif /* __PERF_DATA_H */ Signature · 1 line
2.17.2
Signature · 1 line
- Arnaldo
Re: [PATCH 06/14] perf data: Add perf_data__(create_dir|free_dir) functions
Quoted 10 lines · Arnaldo Carvalho de Melo
Em Sun, Feb 03, 2019 at 04:30:10PM +0100, Jiri Olsa escreveu:
Quoted 6 lines
Adding
perf_data__create_dirto create nr files inside structperf_datapath directory:
intperf_data__create_dir(struct perf_data *data, int nr);and function to free that data:
voidperf_data__free_dir(struct perf_data *data);Can you please rename this to
perf_data__remove_dir(), I think it is more appropriate, just like thatrm_rf()thing.
ok, I will check on the rollback in the new version
jirka
Jiri Olsa
reviewerPATCH 7/14 perf data: Add perf_data__open_dir_data function
In reply to Jiri OlsaAdding perf_data__open_dir_data to open files inside struct perf_data path directory:
static int perf_data__open_dir(struct perf_data *data);
--- tools/perf/util/data.c | 60 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 60 insertions(+)
tools/perf/util/data.c +60 -0
@@ -8,6 +8,8 @@#include <unistd.h> #include <string.h> #include <asm/bug.h> +#include <sys/types.h> +#include <dirent.h> #include "data.h" #include "util.h"
@@ -59,6 +61,64 @@ int perf_data__create_dir(struct perf_data *data, int nr)return ret; } +__maybe_unused +static int perf_data__open_dir (struct perf_data *data) +{ +struct perf_data_file *files =NULL ; +struct dirent *dent; +int ret =-1 ; +DIR *dir; +int nr =0 ; + + dir =opendir (data->path); +if (!dir) +return -EINVAL ; + +while ((dent =readdir (dir)) !=NULL ) { +struct perf_data_file *file; +char path[PATH_MAX ]; +struct stat st; + +snprintf (path,sizeof (path),"%s/%s" , data->path, dent->d_name); +if (stat (path, &st)) +continue ; + +if (!S_ISREG (st.st_mode) ||strncmp (dent->d_name,"data" ,4 )) +continue ; + + ret = -ENOMEM ; + + file =realloc (files, (nr +1 ) *sizeof (*files)); +if (!file) + goto out_err; + + files = file; + file = &files[nr++]; + + file->path =strdup (path); +if (!file->path) + goto out_err; + + ret =open (file->path,O_RDONLY ); +if (ret <0 ) + goto out_err; + + file->fd = ret; + file->size = st.st_size; + } + +if (!files) +return -EINVAL ; + + data->dir.files = files; + data->dir.nr = nr; +return 0 ; + +out_err: +free_dir (files, nr); +return ret; +} +static bool check_pipe (struct perf_data *data) {struct stat st;
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 8/14 perf data: Add directory support
In reply to Jiri OlsaAdding support to have directory as perf.data.
The caller needs to set 'struct perf_data::is_dir flag and the path will be treated as directory.
The 'struct perf_data::file' is initialized and open as 'path/header' file.
Adding check to direcory interface functions to check on is_dir flag.
--- tools/perf/util/data.c | 39 +++++++++++++++++++++++++++++++++++++-- tools/perf/util/data.h | 6 ++++++ 2 files changed, 43 insertions(+), 2 deletions(-)
tools/perf/util/data.c +37 -2
@@ -34,6 +34,9 @@ int perf_data__create_dir(struct perf_data *data, int nr)struct perf_data_file *files =NULL ;int i, ret =-1 ; +if (WARN_ON (!data->is_dir)) +return -EINVAL ; + files =malloc (nr *sizeof (*files));if (!files)return -ENOMEM ;
@@ -61,7 +64,6 @@ int perf_data__create_dir(struct perf_data *data, int nr)return ret; } -__maybe_unused static int perf_data__open_dir (struct perf_data *data) {struct perf_data_file *files =NULL ;
@@ -70,6 +72,9 @@ static int perf_data__open_dir(struct perf_data *data)DIR *dir;int nr =0 ; +if (WARN_ON (!data->is_dir)) +return -EINVAL ; + dir =opendir (data->path);if (!dir)return -EINVAL ;
@@ -159,6 +164,16 @@ static int check_backup(struct perf_data *data)return 0 ; } +static bool is_dir (struct perf_data *data) +{ +struct stat st; + +if (stat (data->path, &st)) +return false; + +return (st.st_mode &S_IFMT ) ==S_IFDIR ; +} +static int open_file_read (struct perf_data *data) {struct stat st;
@@ -240,6 +255,22 @@ static int open_file_dup(struct perf_data *data)return open_file (data); } +static int open_dir (struct perf_data *data) +{ +if (perf_data__is_write (data)) { +if (mkdir (data->path,S_IRWXU ) <0 ) +return -1 ; + }else { +if (perf_data__open_dir (data)) +return -1 ; + } + +if (asprintf (&data->file.path,"%s/header" , data->path) <0 ) +return -ENOMEM ; + +return open_file (data); +} +int perf_data__open (struct perf_data *data) {if (check_pipe (data))
@@ -251,7 +282,11 @@ int perf_data__open(struct perf_data *data)if (check_backup (data))return -1 ; -return open_file_dup (data); +if (perf_data__is_read (data)) + data->is_dir =is_dir (data); + +return perf_data__is_dir (data) ? +open_dir (data) :open_file_dup (data); }void perf_data__close (struct perf_data *data)
tools/perf/util/data.h +6 -0
@@ -19,6 +19,7 @@ struct perf_data {const char *path;struct perf_data_file file;bool is_pipe; +bool is_dir;bool force;enum perf_data_mode mode;
@@ -43,6 +44,11 @@ static inline int perf_data__is_pipe(struct perf_data *data)return data->is_pipe; } +static inline bool perf_data__is_dir (struct perf_data *data) +{ +return data->is_dir; +} +static inline int perf_data__fd (struct perf_data *data) {return data->file.fd;
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 9/14 perf data: Don't store auxtrace index for directory data file
In reply to Jiri OlsaWe can't store auxtrace index when we store to multiple files, because we keep only offset for it, not the file.
The auxtrace data will be processed correctly in the 'pipe' mode.
--- tools/perf/builtin-record.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/builtin-record.c +1 -1
@@ -391,7 +391,7 @@ static int record__process_auxtrace(struct perf_tool *tool,size_t padding;u8 pad[8 ] = {0 }; -if (!perf_data__is_pipe (data)) { +if (!perf_data__is_pipe (data) && !perf_data__is_dir (data)) {off_t file_offset;int fd =perf_data__fd (data);int err;
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 10/14 perf data: Add perf_data__update_dir function
In reply to Jiri OlsaAdding perf_data__update_dir function to update size for every file within the perf.data directory.
--- tools/perf/util/data.c | 20 ++++++++++++++++++++ tools/perf/util/data.h | 1 + 2 files changed, 21 insertions(+)
tools/perf/util/data.c +20 -0
@@ -124,6 +124,26 @@ static int perf_data__open_dir(struct perf_data *data)return ret; } +int perf_data__update_dir (struct perf_data *data) +{ +int i; + +if (WARN_ON (!data->is_dir)) +return -EINVAL ; + +for (i =0 ; i < data->dir.nr; i++) { +struct perf_data_file *file = &data->dir.files[i]; +struct stat st; + +if (fstat (file->fd, &st)) +return -1 ; + + file->size = st.st_size; + } + +return 0 ; +} +static bool check_pipe (struct perf_data *data) {struct stat st;
tools/perf/util/data.h +1 -0
@@ -78,4 +78,5 @@ int perf_data__switch(struct perf_data *data,int perf_data__create_dir (struct perf_data *data,int nr);void perf_data__free_dir (struct perf_data *data); +int perf_data__update_dir (struct perf_data *data);#endif /* __PERF_DATA_H */
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 11/14 perf data: Make perf_data__size to work over directory
In reply to Jiri OlsaMaking perf_data__size to return proper size
for directory data.
--- tools/perf/util/data.c | 17 +++++++++++++++++ tools/perf/util/data.h | 6 +----- 2 files changed, 18 insertions(+), 5 deletions(-)
tools/perf/util/data.c +17 -0
@@ -367,3 +367,20 @@ int perf_data__switch(struct perf_data *data,free (new_filepath);return ret; } + +unsigned long perf_data__size (struct perf_data *data) +{ +u64 size = data->file.size; +int i; + +if (!data->is_dir) +return size; + +for (i =0 ; i < data->dir.nr; i++) { +struct perf_data_file *file = &data->dir.files[i]; + + size += file->size; + } + +return size; +}
tools/perf/util/data.h +1 -5
@@ -54,11 +54,6 @@ static inline int perf_data__fd(struct perf_data *data)return data->file.fd; } -static inline unsigned long perf_data__size (struct perf_data *data) -{ -return data->file.size; -} -int perf_data__open (struct perf_data *data);void perf_data__close (struct perf_data *data); ssize_t perf_data__write(struct perf_data *data,
@@ -79,4 +74,5 @@ int perf_data__switch(struct perf_data *data,int perf_data__create_dir (struct perf_data *data,int nr);void perf_data__free_dir (struct perf_data *data);int perf_data__update_dir (struct perf_data *data); +unsigned long perf_data__size (struct perf_data *data);#endif /* __PERF_DATA_H */
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 12/14 perf session: Add __perf_session__process_dir_events function
In reply to Jiri OlsaAdding __perf_session__process_dir_events function to process events over the directory data.
All directory events are pushed into sessions ordered data and flushed for processing.
--- tools/perf/util/session.c | 86 ++++++++++++++++++++++++++++++++++++++- 1 file changed, 84 insertions(+), 2 deletions(-)
tools/perf/util/session.c +84 -2
@@ -1864,8 +1864,6 @@ reader__process_events(struct reader *rd, struct perf_session *session, file_offset = page_offset; head = rd->data_offset - page_offset; -ui_progress__init_size (prog, data_size,"Processing events..." ); - data_size += rd->data_offset; mmap_size =MMAP_SIZE ;
@@ -1994,6 +1992,87 @@ static int __perf_session__process_events(struct perf_session *session)return err; } +static s64 process_index (struct perf_session *session, +union perf_event *event, +u64 file_offset) +{ +struct perf_evlist *evlist = session->evlist; +u64 timestamp; +s64 ret; + +if (session->header.needs_swap) +event_swap (event,perf_evlist__sample_id_all (evlist)); + +if (event->header.type >=PERF_RECORD_HEADER_MAX ) +return -EINVAL ; + +events_stats__inc (&evlist->stats, event->header.type); + +if (event->header.type >=PERF_RECORD_USER_TYPE_START ) +return perf_session__process_user_event (session, event, file_offset); + + ret =perf_evlist__parse_sample_timestamp (evlist, event, ×tamp); +if (ret) +return ret; + +return ordered_events__queue (&session->ordered_events, event, + timestamp, file_offset); +} + +static int __perf_session__process_dir_events (struct perf_session *session) +{ +struct perf_data *data = session->data; +struct perf_tool *tool = session->tool; +struct reader rd = { + .fd =perf_data__fd (session->data), + .data_size = session->header.data_size, + .data_offset = session->header.data_offset, + .process = process_simple, + }; +int i, ret =0 ; +struct ui_progress prog; +u64 total_size =perf_data__size (session->data); + +perf_tool__fill_defaults (tool); + +ui_progress__init_size (&prog, total_size,"Processing events..." ); + + ret =reader__process_events (&rd, session, &prog); +if (ret) + goto out_err; + +for (i =0 ; i < data->dir.nr ; i++) { +struct perf_data_file *file = &data->dir.files[i]; + +if (file->size ==0 ) +continue ; + + rd = (struct reader ) { + .fd = file->fd, + .data_size = file->size, + .data_offset =0 , + .process = process_index, + }; + + ret =reader__process_events (&rd, session, &prog); +if (ret) + goto out_err; + } + + ret =ordered_events__flush (&session->ordered_events,OE_FLUSH__FINAL ); + +out_err: +if (!tool->no_warn) +perf_session__warn_about_errors (session); + +/* +* We may switching perf.data output, make ordered_events +* reusable. +*/ +ordered_events__reinit (&session->ordered_events); +return ret; +} +int perf_session__process_events (struct perf_session *session) {if (perf_session__register_idle_thread (session) <0 )
@@ -2002,6 +2081,9 @@ int perf_session__process_events(struct perf_session *session)if (perf_data__is_pipe (session->data))return __perf_session__process_pipe_events (session); +if (perf_data__is_dir (session->data)) +return __perf_session__process_dir_events (session); +return __perf_session__process_events (session); }
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 13/14 perf session: Add path to reader object
In reply to Jiri OlsaAdding path to reader object, so we can display file the processing fails for (string in [] brackets).
$ perf report --stdio 0x5e0 [perf.data/data.3] [0xa200]: failed to process type: -1577027574
--- tools/perf/util/session.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
tools/perf/util/session.c +6 -2
@@ -1846,6 +1846,7 @@ struct reader {u64 data_size;u64 data_offset;reader_cb_t process; +char *path; }; static int
@@ -1915,8 +1916,8 @@ reader__process_events(struct reader *rd, struct perf_session *session,if (size <sizeof (struct perf_event_header ) || (skip = rd->process (session, event, file_pos)) <0 ) { -pr_err ("%#" PRIx64" [%#x]: failed to process type: %d\n" , - file_offset + head, event->header.size, +pr_err ("%#" PRIx64" [%s] [%#x]: failed to process type: %d\n" , + file_offset + head, rd->path, event->header.size, event->header.type); err = -EINVAL ; goto out;
@@ -1954,6 +1955,7 @@ static int __perf_session__process_events(struct perf_session *session) .data_size = session->header.data_size, .data_offset = session->header.data_offset, .process = process_simple, + .path = session->data->file.path, };struct ordered_events *oe = &session->ordered_events;struct perf_tool *tool = session->tool;
@@ -2028,6 +2030,7 @@ static int __perf_session__process_dir_events(struct perf_session *session) .data_size = session->header.data_size, .data_offset = session->header.data_offset, .process = process_simple, + .path = session->data->file.path, };int i, ret =0 ;struct ui_progress prog;
@@ -2052,6 +2055,7 @@ static int __perf_session__process_dir_events(struct perf_session *session) .data_size = file->size, .data_offset = 0, .process = process_index, + .path = file->path, }; ret =reader__process_events (&rd, session, &prog);
Signature · 1 line
2.17.2
Jiri Olsa
reviewerPATCH 14/14 perf record: Add --dir option to store data in directory
In reply to Jiri OlsaAdding --dir option to store data in directory. It's next step for multiple threads in record. It's not possible to make directory data via --dir option, like:
$ perf record --dir perf bench sched messaging $ ls -l perf.data total 344 -rw-------. 1 jolsa jolsa 43864 Jan 20 22:26 data.0 -rw-------. 1 jolsa jolsa 30464 Jan 20 22:26 data.1 -rw-------. 1 jolsa jolsa 53816 Jan 20 22:26 data.2 -rw-------. 1 jolsa jolsa 30368 Jan 20 22:26 data.3 -rw-------. 1 jolsa jolsa 40088 Jan 20 22:26 data.4 -rw-------. 1 jolsa jolsa 42592 Jan 20 22:26 data.5 -rw-------. 1 jolsa jolsa 56136 Jan 20 22:26 data.6 -rw-------. 1 jolsa jolsa 25992 Jan 20 22:26 data.7 -rw-------. 1 jolsa jolsa 8832 Jan 20 22:26 header
There's a data file created for every cpu and it's storing data for those cpu maps.
It's possible to transform directory data into standard perf.data file via following inject command:
$ perf inject -o perf.data.file -i perf.data
--- tools/perf/Documentation/perf-record.txt | 3 ++ tools/perf/builtin-record.c | 59 ++++++++++++++++++++++-- tools/perf/util/mmap.h | 23 ++++----- 3 files changed, 70 insertions(+), 15 deletions(-)
tools/perf/Documentation/perf-record.txt +3 -0
@@ -505,6 +505,9 @@ config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'. Implies --tail-synthesize. +--dir:: +Store data into directory with one data file for cpu. + SEE ALSO -------- linkperf:perf-stat[1], linkperf:perf-list[1]
tools/perf/builtin-record.c +55 -4
115 lines not on this page · raw message
tools/perf/util/mmap.h +12 -11
@@ -19,17 +19,18 @@ struct aiocb;* @refcnt - e.g. code using PERF_EVENT_IOC_SET_OUTPUT to share this */ struct perf_mmap { -void *base; -int mask; -int fd; -int cpu; -refcount_t refcnt; -u64 prev; -u64 start; -u64 end; -bool overwrite; -struct auxtrace_mmap auxtrace_mmap; -char event_copy[PERF_SAMPLE_MAX_SIZE ]__aligned(8) ; +void *base; +int mask; +int fd; +int cpu; +refcount_t refcnt; +u64 prev; +u64 start; +u64 end; +bool overwrite; +struct auxtrace_mmap auxtrace_mmap; +struct perf_data_file *file; +char event_copy[PERF_SAMPLE_MAX_SIZE ]__aligned(8) ; #ifdef HAVE_AIO_SUPPORT struct { void **data;
Signature · 1 line
2.17.2
Re: [PATCH 14/14] perf record: Add --dir option to store data in directory
Quoted 38 lines · Jiri Olsa
Adding --dir option to store data in directory. It's next step for multiple threads in record. It's not possible to make directory data via --dir option, like:
$ perf record --dir perf bench sched messaging $ ls -l perf.data total 344 -rw-------. 1 jolsa jolsa 43864 Jan 20 22:26 data.0 -rw-------. 1 jolsa jolsa 30464 Jan 20 22:26 data.1 -rw-------. 1 jolsa jolsa 53816 Jan 20 22:26 data.2 -rw-------. 1 jolsa jolsa 30368 Jan 20 22:26 data.3 -rw-------. 1 jolsa jolsa 40088 Jan 20 22:26 data.4 -rw-------. 1 jolsa jolsa 42592 Jan 20 22:26 data.5 -rw-------. 1 jolsa jolsa 56136 Jan 20 22:26 data.6 -rw-------. 1 jolsa jolsa 25992 Jan 20 22:26 data.7 -rw-------. 1 jolsa jolsa 8832 Jan 20 22:26 headerThere's a data file created for every cpu and it's storing data for those cpu maps.
It's possible to transform directory data into standard perf.data file via following inject command:
$ perf inject -o perf.data.file -i perf.dataSigned-off-by Jiri Olsa <[email protected]>--- tools/perf/Documentation/perf-record.txt | 3 ++ tools/perf/builtin-record.c | 59 ++++++++++++++++++++++-- tools/perf/util/mmap.h | 23 ++++----- 3 files changed, 70 insertions(+), 15 deletions(-)tools/perf/Documentation/perf-record.txt +0
-0
tools/perf/Documentation/perf-record.txt config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'. · line 510
Quoted 6 lines
tools/perf/Documentation/perf-record.txt +3
-0@@ -505,6 +505,9 @@ config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'. Implies --tail-synthesize. +--dir:: +Store data into directory with one data file for cpu. +
Makes sense to mention compatibility with -o option and per-thread buffer mapping.
Quoted 10 lines
tools/perf/Documentation/perf-record.txt +0
-0SEE ALSO -------- linkperf:perf-stat[1], linkperf:perf-list[1]tools/perf/builtin-record.c +0
-0@@ -111,17 +111,21 @@ static bool switch_output_time(struct record *rec)trigger_is_ready (&switch_output_trigger); }
tools/perf/builtin-record.c static bool switch_output_time(struct record *rec) · line 120
Quoted 10 lines
tools/perf/builtin-record.c +5
-2-static int record__write (struct record *rec,struct perf_mmap *map__maybe_unused , +static int record__write (struct record *rec,struct perf_mmap *map,void *bf,size_t size) { -struct perf_data_file *file = &rec->session->data->file; +struct perf_data_file *file = &rec->data.file; + +if (map && map->file) + file = map->file;
For AIO mode per-cpu streaming could be done in parallel because Posix AIO API uses a separate thread for every open data.# or header fd.
tools/perf/builtin-record.c static bool switch_output_time(struct record *rec) · line 128
Quoted 9 lines
tools/perf/builtin-record.c +2
-1if (perf_data_file__write (file, bf, size) <0 ) {pr_err ("failed to write perf data, error: %m\n" );return -1 ; } - rec->bytes_written += size; +if (file == &rec->data.file) + rec->bytes_written += size;
switch-output logic now tracks header file size only? If so, it probably needs to be corrected or simply disabled for --output_dir mode.
Alexey
Quoted 135 lines
tools/perf/builtin-record.c +48
-1if (switch_output_size (rec))trigger_hit (&switch_output_trigger);@@ -563,6 +567,25 @@ static int record__mmap_evlist(struct record *rec,return 0 ; } +static int record__mmap_dir_data (struct record *rec) +{ +struct perf_evlist *evlist = rec->evlist; +struct perf_data *data = &rec->data; +int i, ret, nr = evlist->nr_mmaps; + + ret =perf_data__create_dir (data, nr); +if (ret) +return ret; + +for (i =0 ; i < nr; i++) { +struct perf_mmap *map = &evlist->mmap[i]; + + map->file = &data->dir.files[i]; + } + +return 0 ; +} +static int record__mmap (struct record *rec) {return record__mmap_evlist (rec, rec->evlist);@@ -792,8 +815,12 @@ static int record__mmap_read_evlist(struct record *rec, struct perf_evlist *evli/* * Mark the round finished in case we wrote * at least one event. +* +* No need for round events in directory mode, +* because per-cpu files/maps have sorted data +* from kernel. */ -if (bytes_written != rec->bytes_written) +if (!perf_data__is_dir (&rec->data) && bytes_written != rec->bytes_written) rc =record__write (rec,NULL , &finished_round_event,sizeof (finished_round_event));if (overwrite)@@ -851,6 +878,9 @@ record__finish_output(struct record *rec) rec->session->header.data_size += rec->bytes_written; data->file.size =lseek (perf_data__fd (data),0 ,SEEK_CUR ); +if (perf_data__is_dir (data)) +perf_data__update_dir (data); +if (!rec->no_buildid) {process_buildids (rec);@@ -1170,11 +1200,23 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)if (data->is_pipe && rec->evlist->nr_entries ==1 ) rec->opts.sample_id = true; +if (data->is_pipe &&perf_data__is_dir (data)) { +pr_err ("Directory output is not allowed for pipe output\n" ); + err =-1 ; + goto out_child; + } +if (record__open (rec) !=0 ) { err =-1 ; goto out_child; } +if (perf_data__is_dir (data)) { + err =record__mmap_dir_data (rec); +if (err) + goto out_child; + } + err =bpf__apply_obj_config ();if (err) {char errbuf[BUFSIZ ];@@ -1962,6 +2004,8 @@ static struct option __record_options[] = { &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)", record__aio_parse),#endif +OPT_BOOLEAN (0 ,"dir" , &record.data.is_dir, +"Store data into directory perf.data" ), OPT_END() };@@ -2113,6 +2157,13 @@ int cmd_record(int argc, const char **argv) goto out; } +if (perf_data__is_dir (&rec->data)) { +if (!rec->opts.sample_time) { +pr_err ("Sample timestamp is required for indexing\n" ); + goto out; + } + } +if (rec->opts.target.tid && !rec->opts.no_inherit_set) rec->opts.no_inherit = true;tools/perf/util/mmap.h +12
-11@@ -19,17 +19,18 @@ struct aiocb;* @refcnt - e.g. code using PERF_EVENT_IOC_SET_OUTPUT to share this */ struct perf_mmap { -void *base; -int mask; -int fd; -int cpu; -refcount_t refcnt; -u64 prev; -u64 start; -u64 end; -bool overwrite; -struct auxtrace_mmap auxtrace_mmap; -char event_copy[PERF_SAMPLE_MAX_SIZE ]__aligned(8) ; +void *base; +int mask; +int fd; +int cpu; +refcount_t refcnt; +u64 prev; +u64 start; +u64 end; +bool overwrite; +struct auxtrace_mmap auxtrace_mmap; +struct perf_data_file *file; +char event_copy[PERF_SAMPLE_MAX_SIZE ]__aligned(8) ; #ifdef HAVE_AIO_SUPPORT struct { void **data;
Re: [PATCH 14/14] perf record: Add --dir option to store data in directory
Quoted 41 lines · Alexey Budankov
Quoted 39 lines · Jiri Olsa
Adding --dir option to store data in directory. It's next step for multiple threads in record. It's not possible to make directory data via --dir option, like:
$ perf record --dir perf bench sched messaging $ ls -l perf.data total 344 -rw-------. 1 jolsa jolsa 43864 Jan 20 22:26 data.0 -rw-------. 1 jolsa jolsa 30464 Jan 20 22:26 data.1 -rw-------. 1 jolsa jolsa 53816 Jan 20 22:26 data.2 -rw-------. 1 jolsa jolsa 30368 Jan 20 22:26 data.3 -rw-------. 1 jolsa jolsa 40088 Jan 20 22:26 data.4 -rw-------. 1 jolsa jolsa 42592 Jan 20 22:26 data.5 -rw-------. 1 jolsa jolsa 56136 Jan 20 22:26 data.6 -rw-------. 1 jolsa jolsa 25992 Jan 20 22:26 data.7 -rw-------. 1 jolsa jolsa 8832 Jan 20 22:26 headerThere's a data file created for every cpu and it's storing data for those cpu maps.
It's possible to transform directory data into standard perf.data file via following inject command:
$ perf inject -o perf.data.file -i perf.dataSigned-off-by Jiri Olsa <[email protected]>--- tools/perf/Documentation/perf-record.txt | 3 ++ tools/perf/builtin-record.c | 59 ++++++++++++++++++++++-- tools/perf/util/mmap.h | 23 ++++----- 3 files changed, 70 insertions(+), 15 deletions(-)tools/perf/Documentation/perf-record.txt +0
-0@@ -505,6 +505,9 @@ config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'.
Quoted 7 lines
tools/perf/Documentation/perf-record.txt config terms. For example: 'cycles/overwrite/' and 'instructions/no-overwrite/'. · line 510
Quoted 5 lines
tools/perf/Documentation/perf-record.txt +3
-0Implies --tail-synthesize. +--dir:: +Store data into directory with one data file for cpu. +Makes sense to mention compatibility with -o option and per-thread buffer mapping.
right, will do
Quoted 12 lines
Quoted 12 lines
tools/perf/Documentation/perf-record.txt +0
-0SEE ALSO -------- linkperf:perf-stat[1], linkperf:perf-list[1]tools/perf/builtin-record.c +0
-0@@ -111,17 +111,21 @@ static bool switch_output_time(struct record *rec)trigger_is_ready (&switch_output_trigger); }
Quoted 12 lines
tools/perf/builtin-record.c static bool switch_output_time(struct record *rec) · line 120
Quoted 9 lines
tools/perf/builtin-record.c +5
-2-static int record__write (struct record *rec,struct perf_mmap *map__maybe_unused , +static int record__write (struct record *rec,struct perf_mmap *map,void *bf,size_t size) { -struct perf_data_file *file = &rec->session->data->file; +struct perf_data_file *file = &rec->data.file; + +if (map && map->file) + file = map->file;For AIO mode per-cpu streaming could be done in parallel because Posix AIO API uses a separate thread for every open data.# or header fd.
yes, I was thinking of that and I plan to try that to see if there's some benefit
Quoted 7 lines
Quoted 7 lines
tools/perf/builtin-record.c +0
-0if (perf_data_file__write (file, bf, size) <0 ) {pr_err ("failed to write perf data, error: %m\n" );return -1 ; }
Quoted 7 lines
tools/perf/builtin-record.c static bool switch_output_time(struct record *rec) · line 128
Quoted 3 lines
tools/perf/builtin-record.c +2
-1- rec->bytes_written += size; +if (file == &rec->data.file) + rec->bytes_written += size;switch-output logic now tracks header file size only? If so, it probably needs to be corrected or simply disabled for --output_dir mode.
the rec->bytes_written value is used to track header file size, which is used to fill in the header data section in here record__finish_output, we can't mix other file sizes into it
the switch-output option still needs some changes to work over the directory data
jirka
Hi,
Quoted 9 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
$ perf record --dir perf bench sched messaging
Is it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messaging
Thanks,
Alexey
Quoted 80 lines
$ ls -l perf.data total 344 -rw-------. 1 jolsa jolsa 43864 Jan 20 22:26 data.0 -rw-------. 1 jolsa jolsa 30464 Jan 20 22:26 data.1 -rw-------. 1 jolsa jolsa 53816 Jan 20 22:26 data.2 -rw-------. 1 jolsa jolsa 30368 Jan 20 22:26 data.3 -rw-------. 1 jolsa jolsa 40088 Jan 20 22:26 data.4 -rw-------. 1 jolsa jolsa 42592 Jan 20 22:26 data.5 -rw-------. 1 jolsa jolsa 56136 Jan 20 22:26 data.6 -rw-------. 1 jolsa jolsa 25992 Jan 20 22:26 data.7 -rw-------. 1 jolsa jolsa 8832 Jan 20 22:26 headerThere's a data file created for every cpu and it's storing data for those cpu maps. The report command will read it transparently, sort it and display as single file data.
It's possible to transform directory data into standard perf.data file via simple inject command:
$ perf inject -o perf.data.file -i perf.dataThe old perf fails over the directory data with following message:
$ perf report incompatible file format (rerun with -v to learn more)I'm now testing the record threads support, so I'd like to have some agreement on the directory data support before.
It's also available in here:
git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
perf/dirthanks,
jirka---
Jiri Olsa (14):
perf tools: Makerm_rfto remove single file
perf session: Add process callback to reader object perf data: Move size to structperf_data_file
perf data: Add global path holder
perf data: Makecheck_backupwork over directories perf data: Addperf_data__(create_dir|free_dir)functions perf data: Addperf_data__open_dir_datafunction perf data: Add directory support
perf data: Don't store auxtrace index for directory data file perf data: Addperf_data__update_dirfunction
perf data: Makeperf_data__sizeto work over directory perf session: Add__perf_session__process_dir_eventsfunction perf session: Add path to reader object
perf record: Add --dir option to store data in directorytools/perf/Documentation/perf-record.txt | 3 ++ tools/perf/builtin-annotate.c | 4 +-- tools/perf/builtin-buildid-cache.c | 4 +-- tools/perf/builtin-buildid-list.c | 8 ++--- tools/perf/builtin-c2c.c | 4 +-- tools/perf/builtin-diff.c | 12 +++---- tools/perf/builtin-evlist.c | 4 +-- tools/perf/builtin-inject.c | 10 +++--- tools/perf/builtin-kmem.c | 2 +- tools/perf/builtin-kvm.c | 8 ++--- tools/perf/builtin-lock.c | 8 ++--- tools/perf/builtin-mem.c | 8 ++--- tools/perf/builtin-record.c | 72 +++++++++++++++++++++++++++++++------ tools/perf/builtin-report.c | 6 ++-- tools/perf/builtin-sched.c | 16 ++++----- tools/perf/builtin-script.c | 12 +++---- tools/perf/builtin-stat.c | 6 ++-- tools/perf/builtin-timechart.c | 8 ++--- tools/perf/builtin-trace.c | 8 ++--- tools/perf/util/data-convert-bt.c | 2 +- tools/perf/util/data.c | 231 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/data.h | 24 +++++++++---- tools/perf/util/mmap.h | 23 ++++++------ tools/perf/util/session.c | 117 +++++++++++++++++++++++++++++++++++++++++++++++++++++++----- tools/perf/util/util.c | 10 ++++-- 25 files changed, 479 insertions(+), 131 deletions(-)
Quoted 12 lines · Alexey Budankov
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messaging
yep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output result_1 --dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes run
Total time: 0.177 [sec]
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MB result_1 (7225 samples) ]
[jolsa@krava perf]$ ll result_1/
total 348
-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 header
jirka
Quoted 35 lines · Jiri Olsa
Quoted 12 lines · Alexey Budankov
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messagingyep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output
result_1--dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes runTotal time: 0.177 [sec][ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MBresult_1(7225 samples) ][jolsa@krava perf]$ ll result_1/
total 348
Review
Quoted 9 lines
(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 header
Awesome. What do you think about having it like this:
$ perf record --output result_1.data ... - writes data to a file
$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directory
IMHO, this interface is simpler for a user.
Thanks,
Alexey
Quoted 3 lines
jirka
Quoted 48 lines · Alexey Budankov
Quoted 35 lines · Jiri Olsa
Quoted 12 lines · Alexey Budankov
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messagingyep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output
result_1--dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes runTotal time: 0.177 [sec][ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MBresult_1(7225 samples) ][jolsa@krava perf]$ ll result_1/
total 348Review
Quoted 9 lines
(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
Quoted 6 lines
$ perf record --output result_1.data ... - writes data to a file$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directoryIMHO, this interface is simpler for a user.
yep, seems more convenient.. I'll add it
thanks,
jirka
Quoted 50 lines · Jiri Olsa
Quoted 48 lines · Alexey Budankov
Quoted 35 lines · Jiri Olsa
Quoted 12 lines · Alexey Budankov
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messagingyep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output
result_1--dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes runTotal time: 0.177 [sec][ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MBresult_1(7225 samples) ][jolsa@krava perf]$ ll result_1/
total 348Review
Quoted 9 lines
(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
Quoted 9 lines
Quoted 6 lines
$ perf record --output result_1.data ... - writes data to a file$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directoryIMHO, this interface is simpler for a user.
yep, seems more convenient.. I'll add it
But what happens if you do: perf record -o foo --output_dir foo.d?
Quoted 2 lines
thanks,
jirka
Arnaldo Carvalho de Melo
maintainerEm Mon, Feb 04, 2019 at 10:56:05AM -0800, Stephane Eranian escreveu:
Quoted 53 lines · Jiri Olsa
Quoted 52 lines · Alexey Budankov
Quoted 50 lines · Jiri Olsa
Quoted 35 lines · Alexey Budankov
Quoted 12 lines
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messagingyep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output
result_1--dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes runTotal time: 0.177 [sec][ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MBresult_1(7225 samples) ][jolsa@krava perf]$ ll result_1/
total 348Review
Quoted 9 lines
(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
$ perf record --output result_1.data ... - writes data to a file
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directoryIMHO, this interface is simpler for a user.
yep, seems more convenient.. I'll add it
But what happens if you do: perf record -o foo --output_dir foo.d?
Should fail, i.e. either you use single-file or directory output, I think.
- Arnaldo
Quoted 57 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 04, 2019 at 10:56:05AM -0800, Stephane Eranian escreveu:
Quoted 56 lines · Jiri Olsa
Quoted 55 lines · Alexey Budankov
Quoted 53 lines · Jiri Olsa
Quoted 35 lines · Alexey Budankov
Quoted 12 lines
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messagingyep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output
result_1--dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes runTotal time: 0.177 [sec][ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MBresult_1(7225 samples) ][jolsa@krava perf]$ ll result_1/
total 348Review
Quoted 9 lines
(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
$ perf record --output result_1.data ... - writes data to a file$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directory
Quoted 8 lines
Quoted 5 lines
Quoted 4 lines
Quoted 1 line
IMHO, this interface is simpler for a user.
yep, seems more convenient.. I'll add it
But what happens if you do: perf record -o foo --output_dir foo.d?
Should fail, i.e. either you use single-file or directory output, I think.
I also expect it this way.
Alexey
Quoted 3 lines
- Arnaldo
Quoted 58 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 04, 2019 at 10:56:05AM -0800, Stephane Eranian escreveu:
Quoted 56 lines · Jiri Olsa
Quoted 55 lines · Alexey Budankov
Quoted 53 lines · Jiri Olsa
Quoted 35 lines · Alexey Budankov
Quoted 12 lines
Hi,
Quoted 8 lines · Jiri Olsa
hi,
this patchset adds the --dir option to record command (and all the other record command that overloadcmd_record) that allows the data to be stored in directory with multiple data files.It's next step for multiple threads implementation in record. It's now possible to make directory data via --dir option, like:
Quoted 6 lines
Quoted 1 line
$ perf record --dir perf bench sched messagingIs it possible to name data directory differently from perf.data e.g. using --output option, like this?
$ perf record --output result_1 --dir perf bench sched messagingyep, it's taken into account:
[jolsa@krava perf]$ ./perf record --output
result_1--dir ./perf bench sched messaging Couldn't synthesize bpf events.
# Running 'sched/messaging' benchmark:
# 20 sender and receiver processes per group
# 10 groups == 400 processes runTotal time: 0.177 [sec][ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.316 MBresult_1(7225 samples) ][jolsa@krava perf]$ ll result_1/
total 348Review
Quoted 9 lines
(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
$ perf record --output result_1.data ... - writes data to a file$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directory
Quoted 9 lines
Quoted 5 lines
Quoted 4 lines
Quoted 1 line
IMHO, this interface is simpler for a user.
yep, seems more convenient.. I'll add it
But what happens if you do: perf record -o foo --output_dir foo.d?
Should fail, i.e. either you use single-file or directory output, I think.
Agreed
Quoted 1 line
- Arnaldo
On Mon, Feb 04, 2019 at 12:05:38PM -0800, Stephane Eranian wrote:
SNIP
Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Review
Quoted 12 lines
[jolsa@krava perf]$ ll result_1/
total 348(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
$ perf record --output result_1.data ... - writes data to a file$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directoryIMHO, this interface is simpler for a user.
Quoted 8 lines
Quoted 7 lines
Quoted 3 lines
Quoted 2 lines
yep, seems more convenient.. I'll add it
But what happens if you do: perf record -o foo --output_dir foo.d?
Should fail, i.e. either you use single-file or directory output, I think.
Agreed
ok, will do that
thanks,
jirka
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.
Related to your effort to make perf record multi-threaded, I was wondering how this would impact pipe mode.
Could you explain?
Thanks.
Quoted 40 lines · Jiri Olsa
On Mon, Feb 04, 2019 at 12:05:38PM -0800, Stephane Eranian wrote:
SNIP
Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Review
Quoted 12 lines
[jolsa@krava perf]$ ll result_1/
total 348(quoted diff) +0
-9-rw-------. 1 jolsa jolsa 27624 Feb 4 11:35 data.0 -rw-------. 1 jolsa jolsa 56672 Feb 4 11:35 data.1 -rw-------. 1 jolsa jolsa 30824 Feb 4 11:35 data.2 -rw-------. 1 jolsa jolsa 49136 Feb 4 11:35 data.3 -rw-------. 1 jolsa jolsa 22712 Feb 4 11:35 data.4 -rw-------. 1 jolsa jolsa 53392 Feb 4 11:35 data.5 -rw-------. 1 jolsa jolsa 43352 Feb 4 11:35 data.6 -rw-------. 1 jolsa jolsa 46688 Feb 4 11:35 data.7 -rw-------. 1 jolsa jolsa 9068 Feb 4 11:35 headerAwesome. What do you think about having it like this:
$ perf record --output result_1.data ... - writes data to a file$ perf record --dir result_1 ... - or even $ perf record --output_dir result_1 ... - writes data into a directoryIMHO, this interface is simpler for a user.
Quoted 8 lines
Quoted 7 lines
Quoted 3 lines
Quoted 2 lines
yep, seems more convenient.. I'll add it
But what happens if you do: perf record -o foo --output_dir foo.d?
Should fail, i.e. either you use single-file or directory output, I think.
Agreed
ok, will do that
thanks,
jirka
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
Quoted 4 lines
Related to your effort to make perf record multi-threaded, I was wondering how this would impact pipe mode.
Could you explain?
there's no support for threaded pipe processing at the moment, currently the data are stored in directory:
$ ls -l perf.data total 344 -rw-------. 1 jolsa jolsa 43864 Jan 20 22:26 data.0 -rw-------. 1 jolsa jolsa 30464 Jan 20 22:26 data.1 -rw-------. 1 jolsa jolsa 53816 Jan 20 22:26 data.2 -rw-------. 1 jolsa jolsa 30368 Jan 20 22:26 data.3 -rw-------. 1 jolsa jolsa 40088 Jan 20 22:26 data.4 -rw-------. 1 jolsa jolsa 42592 Jan 20 22:26 data.5 -rw-------. 1 jolsa jolsa 56136 Jan 20 22:26 data.6 -rw-------. 1 jolsa jolsa 25992 Jan 20 22:26 data.7
^^^^ those are raw data files, contains only events with common header
as stored by perf or kernel
-rw-------. 1 jolsa jolsa 8832 Jan 20 22:26 header
^^^^ this one currently holds perf.data file header, describing events
and features
if we switched to pipe mode by default we could omit the header file and find a way to push the data streams through single file, like with a new event describing the data stream.. we can have an option for that so you could do something like:
# perf record --threads --single-file -a ... | perf report -i -
however having single thread storing storing into single file without any other processing is important on record side (for minimal overhead), so I think we should keep creating the directory with data files also for pipe mode to have minimal overhead
jirka
Quoted 19 lines · Jiri Olsa
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.
Quoted 6 lines
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
so following features are not synthesized:
FEAT_OPN(TRACING_DATA, tracing_data, false),
FEAT_OPN(BUILD_ID, build_id, false),
FEAT_OPN(BRANCH_STACK, branch_stack, false),
FEAT_OPN(AUXTRACE, auxtrace, false),
FEAT_OPN(STAT, stat, false),
FEAT_OPN(CACHE, cache, true),I think all could be added and worked around with exception of BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phase
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
jirka
Jiri,
Quoted 28 lines · Jiri Olsa
Quoted 19 lines · Jiri Olsa
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.Quoted 6 lines
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
Quoted 9 lines
so following features are not synthesized:
FEAT_OPN(TRACING_DATA, tracing_data, false), FEAT_OPN(BUILD_ID, build_id, false), FEAT_OPN(BRANCH_STACK, branch_stack, false), FEAT_OPN(AUXTRACE, auxtrace, false), FEAT_OPN(STAT, stat, false), FEAT_OPN(CACHE, cache, true),
What do you need for BRANCH_STACK?
Quoted 4 lines
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phase
Buildids are injected after the fact via perf inject when in pipe mode.
Quoted 4 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
Quoted 5 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
jirka
Quoted 33 lines · Stephane Eranian
Jiri,
Quoted 30 lines · Jiri Olsa
Quoted 19 lines · Jiri Olsa
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.Quoted 6 lines
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
so following features are not synthesized:
Quoted 8 lines
Quoted 7 lines
FEAT_OPN(TRACING_DATA, tracing_data, false), FEAT_OPN(BUILD_ID, build_id, false), FEAT_OPN(BRANCH_STACK, branch_stack, false), FEAT_OPN(AUXTRACE, auxtrace, false), FEAT_OPN(STAT, stat, false), FEAT_OPN(CACHE, cache, true),What do you need for
BRANCH_STACK?
nothing, I think it's just the flag
Quoted 5 lines
Quoted 5 lines
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phase
Quoted 8 lines
Buildids are injected after the fact via perf inject when in pipe mode.
Quoted 4 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
this:
Quoted 5 lines
Quoted 5 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
jirka
jirka
Arnaldo Carvalho de Melo
maintainerEm Mon, Feb 11, 2019 at 07:53:06PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Stephane Eranian
Quoted 7 lines · Jiri Olsa
Quoted 4 lines · Jiri Olsa
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phaseBuildids are injected after the fact via perf inject when in pipe mode.
Quoted 11 lines
Quoted 5 lines
Quoted 3 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
this:
Quoted 3 lines
Quoted 3 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
kernel/bpf/stackmap.c
/* Parse build ID from 64-bit ELF */
static int stack_map_get_build_id_64(void *page_addr,
unsigned char *build_id)
yeah, wasn't aware of that, good thing doing backports, huh? :-)
So do you thing about having a PERF_SAMPLE_BUILDID in sample_type and go and stash that thing in PERF_RECORD_MMAP2?
- Arnaldo
Quoted 24 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 11, 2019 at 07:53:06PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Stephane Eranian
Quoted 7 lines · Jiri Olsa
Quoted 4 lines · Jiri Olsa
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phaseBuildids are injected after the fact via perf inject when in pipe mode.
Quoted 11 lines
Quoted 5 lines
Quoted 3 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
this:
Quoted 3 lines
Quoted 3 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
kernel/bpf/stackmap.c
Quoted 8 lines
/* Parse build ID from 64-bit ELF */
static int stack_map_get_build_id_64(void*page_addr,
unsigned char*build_id)yeah, wasn't aware of that, good thing doing backports, huh? :-)
So do you thing about having a
PERF_SAMPLE_BUILDIDinsample_typeand go and stash that thing inPERF_RECORD_MMAP2?
I thought having new MMAP3 event version with buildid field in it if available and/or enabled by bit in perf_event_attr
jirka
Quoted 26 lines · Jiri Olsa
Quoted 24 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 11, 2019 at 07:53:06PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Stephane Eranian
Quoted 7 lines · Jiri Olsa
Quoted 4 lines · Jiri Olsa
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phaseBuildids are injected after the fact via perf inject when in pipe mode.
Quoted 11 lines
Quoted 5 lines
Quoted 3 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
this:
Quoted 3 lines
Quoted 3 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
kernel/bpf/stackmap.c
Quoted 9 lines
Quoted 8 lines
/* Parse build ID from 64-bit ELF */
static int stack_map_get_build_id_64(void*page_addr,
unsigned char*build_id)yeah, wasn't aware of that, good thing doing backports, huh? :-)
So do you thing about having a
PERF_SAMPLE_BUILDIDinsample_typeand go and stash that thing inPERF_RECORD_MMAP2?
That would be special processing. Normally the sample_type go into each RECORD_SAMPLE.
So I would not recommend this approach.
Quoted 3 lines
I thought having new MMAP3 event version with buildid field in it if available and/or enabled by bit in
perf_event_attr
I think MMAP3 is a cleaner approach, though it adds yet another MMAP event.
Quoted 1 line
jirka
Quoted 36 lines · Stephane Eranian
Quoted 26 lines · Jiri Olsa
Quoted 24 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 11, 2019 at 07:53:06PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Stephane Eranian
Quoted 7 lines · Jiri Olsa
Quoted 4 lines · Jiri Olsa
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phaseBuildids are injected after the fact via perf inject when in pipe mode.
Quoted 11 lines
Quoted 5 lines
Quoted 3 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
this:
Quoted 3 lines
Quoted 3 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
kernel/bpf/stackmap.cQuoted 9 lines
Quoted 8 lines
/* Parse build ID from 64-bit ELF */
static int stack_map_get_build_id_64(void*page_addr,
unsigned char*build_id)yeah, wasn't aware of that, good thing doing backports, huh? :-)
So do you thing about having a
PERF_SAMPLE_BUILDIDinsample_typeand go and stash that thing inPERF_RECORD_MMAP2?
Quoted 8 lines
That would be special processing. Normally the
sample_typego into eachRECORD_SAMPLE.
So I would not recommend this approach.Quoted 3 lines
I thought having new MMAP3 event version with buildid field in it if available and/or enabled by bit in
perf_event_attrI think MMAP3 is a cleaner approach, though it adds yet another MMAP event.
actually I realized this might not help at the end at all
what we do now is that we scan the data (after it's recorded) to get the list of binaries that got touched during the profile and store those in .debug cache based on the build ids
we can't just store ALL of the binaries that the session comes across via mmap events
I can see the build id in mmap helping to resolve the race issue when some binary change during the profile session and we have no idea and report on wrong one.. but I dont see people complaining about this at all
jirka
Arnaldo Carvalho de Melo
maintainerEm Thu, Feb 14, 2019 at 12:34:50PM +0100, Jiri Olsa escreveu:
Quoted 10 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa
Quoted 2 lines
I thought having new MMAP3 event version with buildid field in it if available and/or enabled by bit in
perf_event_attrI think MMAP3 is a cleaner approach, though it adds yet another MMAP event.
actually I realized this might not help at the end at all
what we do now is that we scan the data (after it's recorded)
We wouldn't have to do that scan of the data at the end if we have the build-id in the MMAP3 records.
Quoted 5 lines
to get the list of binaries that got touched during the profile and store those in .debug cache based on the build ids
we can't just store ALL of the binaries that the session comes across via mmap events
We don't have to, we never did that at 'perf record' time, this is something for 'perf archive' to do, see below.
Quoted 4 lines
I can see the build id in mmap helping to resolve the race issue when some binary change during the profile session and we have no idea and report on wrong one.. but I dont see people complaining about this at all
But with the build id in the MMAPs we wouldn't need to scan anything at 'perf record' time, just at 'perf archive' time, where we would scan everything, and as soon as we find a hit, we add that DSO to the list of things we need to put in the tarball.
- Arnaldo
Quoted 29 lines · Arnaldo Carvalho de Melo
Em Thu, Feb 14, 2019 at 12:34:50PM +0100, Jiri Olsa escreveu:
Quoted 10 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa
Quoted 2 lines
I thought having new MMAP3 event version with buildid field in it if available and/or enabled by bit in
perf_event_attrI think MMAP3 is a cleaner approach, though it adds yet another MMAP event.
actually I realized this might not help at the end at all
what we do now is that we scan the data (after it's recorded)
We wouldn't have to do that scan of the data at the end if we have the build-id in the MMAP3 records.
Quoted 5 lines
to get the list of binaries that got touched during the profile and store those in .debug cache based on the build ids
we can't just store ALL of the binaries that the session comes across via mmap events
We don't have to, we never did that at 'perf record' time, this is something for 'perf archive' to do, see below.
Quoted 4 lines
I can see the build id in mmap helping to resolve the race issue when some binary change during the profile session and we have no idea and report on wrong one.. but I dont see people complaining about this at all
Quoted 4 lines
But with the build id in the MMAPs we wouldn't need to scan anything at 'perf record' time, just at 'perf archive' time, where we would scan everything, and as soon as we find a hit, we add that DSO to the list of things we need to put in the tarball.
ok.. it might little change the expected behavour in that you will not have .debug cache populated until you run perf archive.. some profile data might stop report after you reinstall the binary..
on the other hande '.debug' cache would stop growing uncontrolably.. so I think I'd be ok with this
jirka
Arnaldo Carvalho de Melo
maintainerEm Thu, Feb 14, 2019 at 02:26:38PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Arnaldo Carvalho de Melo
Quoted 4 lines
But with the build id in the MMAPs we wouldn't need to scan anything at 'perf record' time, just at 'perf archive' time, where we would scan everything, and as soon as we find a hit, we add that DSO to the list of things we need to put in the tarball.
ok.. it might little change the expected behavour in that you will not have .debug cache populated until you run perf archive.. some profile data might stop report after you reinstall the binary..
Well, nothing prevents us from continuing to, in 'perf record', go thru the perf.data just collected to populate the .debug cache, its just that we would do it just for that, for populating the cache, we wouldn't _have_ to do that for collecting the buildids.
Quoted 2 lines
on the other hand '.debug' cache would stop growing uncontrolably.. so I think I'd be ok with this
That is another problem, and one that has to be solved in a way similar to ccache's --max-size option.
The .debug cache is useful for various workflows, but may get in the way for some others, which is why we have ways to disable it.
For instance, when working on some project it is handy to have copies of binaries saved so that older perf.data files can find a file that was rewritten by the compiler when doing changes in it, etc.
With the .debug cache and if using -g, we can get the older copy of the binary _and_ its sources, annotation for older versions continue to work, etc.
- Arnaldo
Quoted 12 lines · Arnaldo Carvalho de Melo
Em Thu, Feb 14, 2019 at 02:26:38PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Arnaldo Carvalho de Melo
Quoted 4 lines
But with the build id in the MMAPs we wouldn't need to scan anything at 'perf record' time, just at 'perf archive' time, where we would scan everything, and as soon as we find a hit, we add that DSO to the list of things we need to put in the tarball.
ok.. it might little change the expected behavour in that you will not have .debug cache populated until you run perf archive.. some profile data might stop report after you reinstall the binary..
Today perf record does collect the buildids once monitor is completed, it does one
pass over the perf.data file looking for MMAP records, or at least in the version I am
more familiar with.
Quoted 4 lines
Well, nothing prevents us from continuing to, in 'perf record', go thru the perf.data just collected to populate the .debug cache, its just that we would do it just for that, for populating the cache, we wouldn't _have_ to do that for collecting the buildids.
Sure, for compatibility reasons.
In pipe mode, this would also avoid the need for perf inject -b -i -, which is a win.
perf archive is useful if you do not have a way to locate the binary using only the buildid on the analysis machine.
Quoted 4 lines
Quoted 3 lines
on the other hand '.debug' cache would stop growing uncontrolably.. so I think I'd be ok with this
I agree!
Quoted 6 lines
That is another problem, and one that has to be solved in a way similar to ccache's --max-size option.
The .debug cache is useful for various workflows, but may get in the way for some others, which is why we have ways to disable it.
Correct.
Quoted 9 lines
For instance, when working on some project it is handy to have copies of binaries saved so that older perf.data files can find a file that was rewritten by the compiler when doing changes in it, etc.
With the .debug cache and if using -g, we can get the older copy of the binary _and_ its sources, annotation for older versions continue to work, etc.
- Arnaldo
A message that is not in the archive (sent off-list or not yet received): CA+JHD90ssKi3CJ7yfCFTkrS8xwUsZhvd0t7cSCy1MF7TJ2XLYw@mail.gmail.com
Quoted 17 lines · Arnaldo Carvalho de Melo
Quoted 15 lines · On Thu, Feb 14, 2019, 6:31 PM Stephane Eranian <[email protected] wrote:
Quoted 12 lines · Arnaldo Carvalho de Melo
Em Thu, Feb 14, 2019 at 02:26:38PM +0100, Jiri Olsa escreveu:
Quoted 9 lines · Arnaldo Carvalho de Melo
Quoted 4 lines
But with the build id in the MMAPs we wouldn't need to scan anything at 'perf record' time, just at 'perf archive' time, where we would scan everything, and as soon as we find a hit, we add that DSO to the list of things we need to put in the tarball.
ok.. it might little change the expected behavour in that you will not have .debug cache populated until you run perf archive.. some profile data might stop report after you reinstall the binary..
Quoted 9 lines
Quoted 5 lines
Today perf record does collect the buildids once monitor is completed, it does one
pass over the perf.data file looking for MMAP records, or at least in the version I am
more familiar with.It does look for the MMAP records, how else would it find the DSO pathnames?
I was not talking about the synthesize_mmap() code. I was talking about process_buildids().
Quoted 40 lines
- Arnaldo
Sent from smartphone
Quoted 36 lines
Quoted 6 lines
Well, nothing prevents us from continuing to, in 'perf record', go thru the perf.data just collected to populate the .debug cache, its just that we would do it just for that, for populating the cache, we wouldn't _have_ to do that for collecting the buildids.
Sure, for compatibility reasons.
In pipe mode, this would also avoid the need for perf inject -b -i -, which is a win.perf archive is useful if you do not have a way to locate the binary using only the buildid on the analysis machine.
Quoted 4 lines
Quoted 3 lines
on the other hand '.debug' cache would stop growing uncontrolably.. so I think I'd be ok with this
I agree!
Quoted 6 lines
That is another problem, and one that has to be solved in a way similar to ccache's --max-size option.
The .debug cache is useful for various workflows, but may get in the way for some others, which is why we have ways to disable it.
Correct.
Quoted 9 lines
For instance, when working on some project it is handy to have copies of binaries saved so that older perf.data files can find a file that was rewritten by the compiler when doing changes in it, etc.
With the .debug cache and if using -g, we can get the older copy of the binary _and_ its sources, annotation for older versions continue to work, etc.
- Arnaldo
Arnaldo Carvalho de Melo
maintainerEm Mon, Feb 11, 2019 at 10:34:16AM -0800, Stephane Eranian escreveu:
Quoted 46 lines
Jiri,
Quoted 28 lines · Jiri Olsa
Quoted 19 lines · Jiri Olsa
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.Quoted 6 lines
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
Quoted 9 lines
so following features are not synthesized:
FEAT_OPN(TRACING_DATA, tracing_data, false), FEAT_OPN(BUILD_ID, build_id, false), FEAT_OPN(BRANCH_STACK, branch_stack, false), FEAT_OPN(AUXTRACE, auxtrace, false), FEAT_OPN(STAT, stat, false), FEAT_OPN(CACHE, cache, true),What do you need for
BRANCH_STACK?Quoted 4 lines
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phase
Quoted 7 lines
Buildids are injected after the fact via perf inject when in pipe mode.
Quoted 4 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Sure, but what do you propose?
this keeps resurfacing, the idea is to have the building go together with the PERF_RECORD_MMAP3 event, i.e. as part of setting up an executable mapping the loader would get the buildid and ask the kernel to keep it aroung, then when a PERF_RECORD_MMAP needs to be issued, it can include the build id, so tooling will not need to get it.
Alternatively, we would have a separate thread to process PERF_RECORD_MMAP events, and as soon as it gets one from the kernel, augment it straight away with the build-id it reads from the ELF file, i.e. no need to have the kernel provide it, do it just like we do with PERF_RECORD_BPF_EVENT, which reminds me Song probably already posted thise bits...
Quoted 5 lines
Quoted 5 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
jirka
Signature · 1 line
- Arnaldo
Arnaldo,
Quoted 54 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 11, 2019 at 10:34:16AM -0800, Stephane Eranian escreveu:
Quoted 52 lines
Jiri,
Quoted 28 lines · Jiri Olsa
Quoted 19 lines · Jiri Olsa
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.Quoted 6 lines
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
Quoted 9 lines
so following features are not synthesized:
FEAT_OPN(TRACING_DATA, tracing_data, false), FEAT_OPN(BUILD_ID, build_id, false), FEAT_OPN(BRANCH_STACK, branch_stack, false), FEAT_OPN(AUXTRACE, auxtrace, false), FEAT_OPN(STAT, stat, false), FEAT_OPN(CACHE, cache, true),What do you need for
BRANCH_STACK?Quoted 4 lines
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phaseBuildids are injected after the fact via perf inject when in pipe mode.
Quoted 4 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Quoted 8 lines
Quoted 1 line
Sure, but what do you propose?
this keeps resurfacing, the idea is to have the building go together with the
PERF_RECORD_MMAP3event, i.e. as part of setting up an executable mapping the loader would get the buildid and ask the kernel to keep it aroung, then when aPERF_RECORD_MMAPneeds to be issued, it can include the build id, so tooling will not need to get it.
And how would the dynamic loader (ld.so) communicate the buildid to the kernel? How would that work for statically linked binaries.
I think you're say the kernel would parse the ELF header looking for that note section
and extract the buildid from there. Is that what you are proposing?
Quoted 7 lines
Alternatively, we would have a separate thread to process
PERF_RECORD_MMAPevents, and as soon as it gets one from the kernel, augment it straight away with the build-id it reads from the ELF file, i.e. no need to have the kernel provide it, do it just like we do withPERF_RECORD_BPF_EVENT, which reminds me Song probably already posted thise bits...
But that would not work in pipe mode, wouldn't it? Unless that thread intercepts everything pushed to the pipe looking for MMAP records.
Quoted 9 lines
Quoted 5 lines
Quoted 5 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
jirka
--
- Arnaldo
Quoted 70 lines · Stephane Eranian
Arnaldo,
Quoted 54 lines · Arnaldo Carvalho de Melo
Em Mon, Feb 11, 2019 at 10:34:16AM -0800, Stephane Eranian escreveu:
Quoted 52 lines
Jiri,
Quoted 28 lines · Jiri Olsa
Quoted 19 lines · Jiri Olsa
Quoted 17 lines · Stephane Eranian
Jiri,
While you're looking at the output format, I think it would be good time to simplify the code handling perf.data file.
Today, perf record can emit in two formats: file mode or pipe mode. This adds complexity in the code and
is error prone as the file mode path is tested more than the pipe mode path. We have run into multiple issues with
the pipe mode in recent years. There is no real reason why we need to maintain two formats. If I recall, the pipe format
was introduced because on pipes you cannot lseek to update the headers and therefore some of the information present as tables updated on the fly needed to be generated as pseudo records by the tool. I believe that the pipe format covers all the needs and could supersede the file mode format. That would simplify code in perf record and eliminate the risk of errors when new headers are introduced.Quoted 6 lines
yep, I think we have almost all the features covered for pipe mode, and we have all necessary events to describe events features
so with some effort we could switch off the superfluos file header and use only events to describe events ;-) make sense, I'll check on it
Quoted 9 lines
so following features are not synthesized:
FEAT_OPN(TRACING_DATA, tracing_data, false), FEAT_OPN(BUILD_ID, build_id, false), FEAT_OPN(BRANCH_STACK, branch_stack, false), FEAT_OPN(AUXTRACE, auxtrace, false), FEAT_OPN(STAT, stat, false), FEAT_OPN(CACHE, cache, true),What do you need for
BRANCH_STACK?Quoted 4 lines
I think all could be added and worked around with exception of
BUILD_ID, which we store at the end (after processing all data) and we need it early in the report phaseBuildids are injected after the fact via perf inject when in pipe mode.
Quoted 4 lines
maybe it's time to re-think that buildid -> mmap event association again, because it's pain in current implementation as well
Quoted 8 lines
Quoted 1 line
Sure, but what do you propose?
this keeps resurfacing, the idea is to have the building go together with the
PERF_RECORD_MMAP3event, i.e. as part of setting up an executable mapping the loader would get the buildid and ask the kernel to keep it aroung, then when aPERF_RECORD_MMAPneeds to be issued, it can include the build id, so tooling will not need to get it.
Quoted 5 lines
And how would the dynamic loader (ld.so) communicate the buildid to the kernel? How would that work for statically linked binaries.
I think you're say the kernel would parse the ELF header looking for that note section
and extract the buildid from there. Is that what you are proposing?
We have kernel parses ELF header for BUILD-ID in BPF side. You can find the code in stack_map_get_build_id_offset() and functions called by it.
Quoted 11 lines
Quoted 8 lines
Alternatively, we would have a separate thread to process
PERF_RECORD_MMAPevents, and as soon as it gets one from the kernel, augment it straight away with the build-id it reads from the ELF file, i.e. no need to have the kernel provide it, do it just like we do withPERF_RECORD_BPF_EVENT, which reminds me Song probably already posted thise bits...But that would not work in pipe mode, wouldn't it? Unless that thread intercepts everything pushed to the pipe looking for MMAP records.
For PERF_RECORD_BPF_EVENT, I am adding a separate thread, which only listen to PERF_RECORD_BPF_EVENT with watermark of 1. This means, each PERF_RECORD_BPF_EVENT is sent to two ring buffers. One of them got written to the pipe, the other is only processed by the listening thread. Please see https://patchwork.ozlabs.org/patch/1039091/ for details.
Thanks,
Song
Quoted 10 lines
Quoted 10 lines
Quoted 6 lines
Quoted 6 lines
looks like bpf code is actualy getting build ids and storing it for the callchains in kernel.. we can check if we can do something similar for mmap events
jirka
--
- Arnaldo