perf top: Get rid of *_threaded() functions
Namhyung Kim
maintainerPATCH 1/3 perf top: Get rid of *_threaded() functions
From: Namhyung Kim <[email protected]>
Those _threaded() functions are needed to make hist tree handling thread-safe, but AFAICS the only thing it does is forcing it to use the intermediate 'collapsed' tree. It can be acheived by setting sort__need_collapse to 1 in cmd_top() so no need to keep those _threaded() variants.
--- tools/perf/builtin-top.c | 23 +++++++++++++---------- tools/perf/util/hist.c | 44 ++++++-------------------------------------- tools/perf/util/hist.h | 4 ---- 3 files changed, 19 insertions(+), 52 deletions(-)
tools/perf/builtin-top.c +13 -10
@@ -284,11 +284,11 @@ static void perf_top__print_sym_table(struct perf_top *top)return ; } -hists__collapse_resort_threaded (&top->sym_evsel->hists); -hists__output_resort_threaded (&top->sym_evsel->hists); -hists__decay_entries_threaded (&top->sym_evsel->hists, - top->hide_user_symbols, - top->hide_kernel_symbols); +hists__collapse_resort (&top->sym_evsel->hists); +hists__output_resort (&top->sym_evsel->hists); +hists__decay_entries (&top->sym_evsel->hists, + top->hide_user_symbols, + top->hide_kernel_symbols);hists__output_recalc_col_len (&top->sym_evsel->hists, top->print_entries - printed);putchar ('\n' );
@@ -549,11 +549,11 @@ static void perf_top__sort_new_samples(void *arg)if (t->evlist->selected !=NULL ) t->sym_evsel = t->evlist->selected; -hists__collapse_resort_threaded (&t->sym_evsel->hists); -hists__output_resort_threaded (&t->sym_evsel->hists); -hists__decay_entries_threaded (&t->sym_evsel->hists, - t->hide_user_symbols, - t->hide_kernel_symbols); +hists__collapse_resort (&t->sym_evsel->hists); +hists__output_resort (&t->sym_evsel->hists); +hists__decay_entries (&t->sym_evsel->hists, + t->hide_user_symbols, + t->hide_kernel_symbols); }static void *display_thread_tui (void *arg)
@@ -1126,6 +1126,9 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)if (setup_sorting () <0 )usage_with_options (top_usage, options); +/* display thread wants entries to be collapsed in a different tree */ + sort__need_collapse =1 ; +if (top.use_stdio) use_browser =0 ;else if (top.use_tui)
tools/perf/util/hist.c +6 -38
97 lines not on this page · raw message
tools/perf/util/hist.h +0 -4
@@ -104,13 +104,9 @@ struct hist_entry *__hists__add_mem_entry(struct hists *self, u64 weight);void hists__output_resort (struct hists *self); -void hists__output_resort_threaded (struct hists *hists);void hists__collapse_resort (struct hists *self); -void hists__collapse_resort_threaded (struct hists *hists);void hists__decay_entries (struct hists *hists,bool zap_user,bool zap_kernel); -void hists__decay_entries_threaded (struct hists *hists,bool zap_user, -bool zap_kernel);void hists__output_recalc_col_len (struct hists *hists,int max_rows);void hists__inc_nr_entries (struct hists *hists,struct hist_entry *h);
Signature · 1 line
1.7.11.7
Namhyung Kim
maintainerPATCH 2/3 perf hists: Move locking to its call-sites
From: Namhyung Kim <[email protected]>
It's a preparation patch to eliminate unneeded locking in the perf report path.
--- tools/perf/builtin-report.c | 26 ++++++++++++++------------ tools/perf/builtin-top.c | 3 +++ tools/perf/util/hist.c | 6 +----- 3 files changed, 18 insertions(+), 17 deletions(-)
tools/perf/builtin-report.c +14 -12
@@ -297,6 +297,7 @@ static int process_sample_event(struct perf_tool *tool, { struct perf_report *rep = container_of(tool, struct perf_report, tool);struct addr_location al; +int ret;if (perf_event__preprocess_sample (event, machine, &al, sample, rep->annotate_init) <0 ) {
@@ -311,28 +312,29 @@ static int process_sample_event(struct perf_tool *tool,if (rep->cpu_list && !test_bit (sample->cpu, rep->cpu_bitmap))return 0 ; +pthread_mutex_lock (&evsel->hists.lock); +if (sort__mode ==SORT_MODE__BRANCH ) { -if (perf_report__add_branch_hist_entry (tool, &al, sample, - evsel, machine)) { + ret =perf_report__add_branch_hist_entry (tool, &al, sample, + evsel, machine); +if (ret <0 )pr_debug ("problem adding lbr entry, skipping event\n" ); -return -1 ; - } }else if (rep->mem_mode ==1 ) { -if (perf_report__add_mem_hist_entry (tool, &al, sample, - evsel, machine, event)) { + ret =perf_report__add_mem_hist_entry (tool, &al, sample, + evsel, machine, event); +if (ret <0 )pr_debug ("problem adding mem entry, skipping event\n" ); -return -1 ; - } }else {if (al.map !=NULL ) al.map->dso->hit =1 ; -if (perf_evsel__add_hist_entry (evsel, &al, sample, machine)) { + ret =perf_evsel__add_hist_entry (evsel, &al, sample, machine); +if (ret <0 )pr_debug ("problem incrementing symbol period, skipping event\n" ); -return -1 ; - } } -return 0 ; +pthread_mutex_unlock (&evsel->hists.lock); + +return ret; } static int process_read_event(struct perf_tool *tool,
tools/perf/builtin-top.c +3 -0
@@ -245,8 +245,11 @@ static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel, {struct hist_entry *he; +pthread_mutex_lock (&evsel->hists.lock); he =__hists__add_entry (&evsel->hists, al,NULL , sample->period, sample->weight); +pthread_mutex_unlock (&evsel->hists.lock); +if (he ==NULL )return NULL ;
tools/perf/util/hist.c +1 -5
@@ -347,8 +347,6 @@ static struct hist_entry *add_hist_entry(struct hists *hists,struct hist_entry *he;int cmp; -pthread_mutex_lock (&hists->lock); - p = &hists->entries_in->rb_node;while (*p !=NULL ) {
@@ -394,14 +392,12 @@ static struct hist_entry *add_hist_entry(struct hists *hists, he =hist_entry__new (entry);if (!he) - goto out_unlock; +return NULL ;rb_link_node (&he->rb_node_in, parent, p);rb_insert_color (&he->rb_node_in, hists->entries_in); out:hist_entry__add_cpumode_period (he, al->cpumode, period); -out_unlock: -pthread_mutex_unlock (&hists->lock);return he; }
Signature · 1 line
1.7.11.7
Namhyung Kim
maintainerPATCH 3/3 perf report: Don't be bothered with locking when adding hist entries
In reply to Namhyung KimFrom: Namhyung Kim <[email protected]>
The perf report is single-threaded, so no need to grab a lock. Although the fast path of pthread_mutex_[un]lock() is very fast, there's ~3% gain by eliminating it when we have huge sample data.
$ perf record -a -F 100000 -o perf.data.bench -- perf bench sched all $ perf record -e cycles:upp -o perf.data.before -- \
Quoted 1 line
perf report -i perf.data.bench --stdio >
/dev/null
... apply this patch ...
$ perf record -e cycles:upp -o perf.data.after -- \
Quoted 1 line
perf report -i perf.data.bench --stdio >
/dev/null
$ perf diff perf.data.{before,after} | grep pthread
+0.02% libpthread-2.15.so [.] _pthread_cleanup_push_defer
+0.02% libpthread-2.15.so [.] _pthread_cleanup_pop_restore
0.05% -0.05% perf [.] pthread_mutex_unlock@plt
0.05% -0.05% perf [.] pthread_mutex_lock@plt
1.01% -1.01% libpthread-2.15.so [.] pthread_mutex_lock
1.68% -1.68% libpthread-2.15.so [.] __pthread_mutex_unlock_usercnt
0.05% -0.05% libpthread-2.15.so [.] pthread_mutex_unlock--- tools/perf/builtin-report.c | 4 ---- 1 file changed, 4 deletions(-)
tools/perf/builtin-report.c +0 -4
@@ -312,8 +312,6 @@ static int process_sample_event(struct perf_tool *tool,if (rep->cpu_list && !test_bit (sample->cpu, rep->cpu_bitmap))return 0 ; -pthread_mutex_lock (&evsel->hists.lock); -if (sort__mode ==SORT_MODE__BRANCH ) { ret =perf_report__add_branch_hist_entry (tool, &al, sample, evsel, machine);
@@ -332,8 +330,6 @@ static int process_sample_event(struct perf_tool *tool,if (ret <0 )pr_debug ("problem incrementing symbol period, skipping event\n" ); } -pthread_mutex_unlock (&evsel->hists.lock); -return ret; }
Signature · 1 line
1.7.11.7