Re: [PATCH] perf trace: Implement syscall summary in BPF
From: Namhyung Kim
Date: Sat Mar 15 2025 - 20:23:43 EST
Hello Howard,
On Thu, Mar 06, 2025 at 05:57:33PM -0800, Howard Chu wrote:
> Hello Namhyung,
>
> Sorry for replying this crazy late, flu got me good last week...
I'm also late. :) Hope you are well now.
>
> On Fri, Feb 21, 2025 at 10:32 AM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> >
> > When --summary-only option is used, it doesn't need (augmented) arguments
> > of syscalls. Let's skip the augmentation and load another small BPF
> > program to collect the statistics in the kernel instead of copying the
> > data to the ring-buffer to calculate the stats in userspace. This will
> > be much more light-weight than the existing approach and remove any lost
> > events.
>
> Yep, awesome idea, 100% agree. That's how the summary mode should be
> written, so cool :)
>
> >
> > For simplicity, it's only activated with --summary-mode=total in system-
> > wide mode. And it also skips to calculate stddev as doing it atomically
> > would add more overheads (i.e. requiring a spinlock). It can be extended
> > to cover more use cases later, if needed.
> >
> > Cc: Howard Chu <howardchu95@xxxxxxxxx>
> > Signed-off-by: Namhyung Kim <namhyung@xxxxxxxxxx>
> > ---
> > tools/perf/Makefile.perf | 2 +-
> > tools/perf/builtin-trace.c | 41 ++++--
> > tools/perf/util/Build | 1 +
> > tools/perf/util/bpf-trace-summary.c | 117 ++++++++++++++++++
> > .../perf/util/bpf_skel/syscall_summary.bpf.c | 109 ++++++++++++++++
> > tools/perf/util/bpf_skel/syscall_summary.h | 18 +++
> > tools/perf/util/trace.h | 31 +++++
> > 7 files changed, 311 insertions(+), 8 deletions(-)
> > create mode 100644 tools/perf/util/bpf-trace-summary.c
> > create mode 100644 tools/perf/util/bpf_skel/syscall_summary.bpf.c
> > create mode 100644 tools/perf/util/bpf_skel/syscall_summary.h
> > create mode 100644 tools/perf/util/trace.h
> >
> > diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
> > index 55d6ce9ea52fb2a5..3b68a57682ca991d 100644
> > --- a/tools/perf/Makefile.perf
> > +++ b/tools/perf/Makefile.perf
> > @@ -1198,7 +1198,7 @@ SKELETONS += $(SKEL_OUT)/bperf_leader.skel.h $(SKEL_OUT)/bperf_follower.skel.h
> > SKELETONS += $(SKEL_OUT)/bperf_cgroup.skel.h $(SKEL_OUT)/func_latency.skel.h
> > SKELETONS += $(SKEL_OUT)/off_cpu.skel.h $(SKEL_OUT)/lock_contention.skel.h
> > SKELETONS += $(SKEL_OUT)/kwork_trace.skel.h $(SKEL_OUT)/sample_filter.skel.h
> > -SKELETONS += $(SKEL_OUT)/kwork_top.skel.h
> > +SKELETONS += $(SKEL_OUT)/kwork_top.skel.h $(SKEL_OUT)/syscall_summary.skel.h
> > SKELETONS += $(SKEL_OUT)/bench_uprobe.skel.h
> > SKELETONS += $(SKEL_OUT)/augmented_raw_syscalls.skel.h
> >
> > diff --git a/tools/perf/builtin-trace.c b/tools/perf/builtin-trace.c
> > index f55a8a6481f27f99..1c7af7583bc560c4 100644
> > --- a/tools/perf/builtin-trace.c
> > +++ b/tools/perf/builtin-trace.c
> > @@ -55,6 +55,7 @@
> > #include "util/thread_map.h"
> > #include "util/stat.h"
> > #include "util/tool.h"
> > +#include "util/trace.h"
> > #include "util/util.h"
> > #include "trace/beauty/beauty.h"
> > #include "trace-event.h"
> > @@ -222,6 +223,7 @@ struct trace {
> > bool force;
> > bool vfs_getname;
> > bool force_btf;
> > + bool summary_bpf;
> > int trace_pgfaults;
> > char *perfconfig_events;
> > struct {
> > @@ -4276,6 +4278,13 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
> >
> > trace->live = true;
> >
> > + if (trace->summary_bpf) {
> > + if (trace_prepare_bpf_summary() < 0)
> > + goto out_delete_evlist;
> > +
> > + goto create_maps;
> > + }
> > +
> > if (!trace->raw_augmented_syscalls) {
> > if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
> > goto out_error_raw_syscalls;
> > @@ -4334,6 +4343,7 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
> > if (trace->cgroup)
> > evlist__set_default_cgroup(trace->evlist, trace->cgroup);
> >
> > +create_maps:
> > err = evlist__create_maps(evlist, &trace->opts.target);
> > if (err < 0) {
> > fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
> > @@ -4426,9 +4436,11 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
> > if (err < 0)
> > goto out_error_apply_filters;
> >
> > - err = evlist__mmap(evlist, trace->opts.mmap_pages);
> > - if (err < 0)
> > - goto out_error_mmap;
> > + if (!trace->summary_bpf) {
> > + err = evlist__mmap(evlist, trace->opts.mmap_pages);
> > + if (err < 0)
> > + goto out_error_mmap;
> > + }
> >
> > if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
> > evlist__enable(evlist);
>
> I think if summary_bpf is true we don't need to enable the evlist
> either, but we can leave it here.
Probably. I think I just wanted to make minimal changes to the existing
code if possible.
>
> > @@ -4441,6 +4453,9 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
> > evlist__enable(evlist);
> > }
> >
> > + if (trace->summary_bpf)
> > + trace_start_bpf_summary();
> > +
> > trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
> > perf_thread_map__nr(evlist->core.threads) > 1 ||
> > evlist__first(evlist)->core.attr.inherit;
> > @@ -4508,12 +4523,17 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
> >
> > evlist__disable(evlist);
> >
> > + if (trace->summary_bpf)
> > + trace_end_bpf_summary();
> > +
> > if (trace->sort_events)
> > ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
> >
> > if (!err) {
> > if (trace->summary) {
> > - if (trace->summary_mode == SUMMARY__BY_TOTAL)
> > + if (trace->summary_bpf)
> > + trace_print_bpf_summary(trace->sctbl, trace->output);
> > + else if (trace->summary_mode == SUMMARY__BY_TOTAL)
> > trace__fprintf_total_summary(trace, trace->output);
> > else
> > trace__fprintf_thread_summary(trace, trace->output);
> > @@ -4529,6 +4549,7 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
> > }
> >
> > out_delete_evlist:
> > + trace_cleanup_bpf_summary();
> > delete_syscall_stats(trace->syscall_stats);
> > trace__symbols__exit(trace);
> > evlist__free_syscall_tp_fields(evlist);
> > @@ -5446,6 +5467,15 @@ int cmd_trace(int argc, const char **argv)
> > goto skip_augmentation;
> > }
> >
> > + if (!argc && target__none(&trace.opts.target))
> > + trace.opts.target.system_wide = true;
> > +
> > + if (trace.summary_only && trace.opts.target.system_wide &&
> > + trace.summary_mode == SUMMARY__BY_TOTAL && !trace.trace_pgfaults) {
>
> I feel like trace_pgfaults can coexist with the BPF summary, but I'm
> fine with the current logic.
>
> BTW, this part of the code can be confusing; You covered a case where:
> sudo ./perf trace -s --summary-mode=total -F all -a --syscall
> is handled properly. Because normally when doing
> sudo ./perf trace -s --summary-mode=total -F all -a
> without the --syscall option, perf trace won't even get to this point,
> it will skip augmentation from here:
> if (!trace.trace_syscalls)
> goto skip_augmentation;
>
> So, good work!
I feel like I'd add an option to turn it on explicitly and support
per-task summary as well.
Maybe something like --bpf-summary?
>
> > + trace.summary_bpf = true;
> > + goto skip_augmentation;
> > + }
> > +
> > trace.skel = augmented_raw_syscalls_bpf__open();
> > if (!trace.skel) {
> > pr_debug("Failed to open augmented syscalls BPF skeleton");
> > @@ -5649,9 +5679,6 @@ int cmd_trace(int argc, const char **argv)
> > goto out_close;
> > }
> >
> > - if (!argc && target__none(&trace.opts.target))
> > - trace.opts.target.system_wide = true;
> > -
> > if (input_name)
> > err = trace__replay(&trace);
> > else
> > diff --git a/tools/perf/util/Build b/tools/perf/util/Build
> > index 034a6603d5a8e8b0..ba4201a6f3c69753 100644
> > --- a/tools/perf/util/Build
> > +++ b/tools/perf/util/Build
> > @@ -171,6 +171,7 @@ perf-util-$(CONFIG_PERF_BPF_SKEL) += bpf_off_cpu.o
> > perf-util-$(CONFIG_PERF_BPF_SKEL) += bpf-filter.o
> > perf-util-$(CONFIG_PERF_BPF_SKEL) += bpf-filter-flex.o
> > perf-util-$(CONFIG_PERF_BPF_SKEL) += bpf-filter-bison.o
> > +perf-util-$(CONFIG_PERF_BPF_SKEL) += bpf-trace-summary.o
> > perf-util-$(CONFIG_PERF_BPF_SKEL) += btf.o
> >
> > ifeq ($(CONFIG_LIBTRACEEVENT),y)
> > diff --git a/tools/perf/util/bpf-trace-summary.c b/tools/perf/util/bpf-trace-summary.c
> > new file mode 100644
> > index 0000000000000000..7e8b1c9b3faeee4f
> > --- /dev/null
> > +++ b/tools/perf/util/bpf-trace-summary.c
> > @@ -0,0 +1,117 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +#include <inttypes.h>
> > +#include <stdio.h>
> > +#include <stdlib.h>
> > +
> > +#include "syscalltbl.h"
> > +#include "util/trace.h"
> > +#include "util/util.h"
> > +#include <bpf/bpf.h>
> > +#include <linux/time64.h>
> > +
> > +#include "bpf_skel/syscall_summary.h"
> > +#include "bpf_skel/syscall_summary.skel.h"
> > +
> > +
> > +static struct syscall_summary_bpf *skel;
> > +
> > +int trace_prepare_bpf_summary(void)
> > +{
> > + skel = syscall_summary_bpf__open_and_load();
> > + if (skel == NULL) {
> > + fprintf(stderr, "failed to load syscall summary bpf skeleton\n");
> > + return -1;
> > + }
> > +
> > + if (syscall_summary_bpf__attach(skel) < 0) {
> > + fprintf(stderr, "failed to attach syscall summary bpf skeleton\n");
> > + return -1;
> > + }
> > +
> > + return 0;
> > +}
> > +
> > +void trace_start_bpf_summary(void)
> > +{
> > + skel->bss->enabled = 1;
> > +}
> > +
> > +void trace_end_bpf_summary(void)
> > +{
> > + skel->bss->enabled = 0;
> > +}
> > +
> > +struct syscall_data {
> > + int syscall_nr;
> > + struct syscall_stats stats;
> > +};
> > +
> > +static int datacmp(const void *a, const void *b)
> > +{
> > + const struct syscall_data *sa = a;
> > + const struct syscall_data *sb = b;
> > +
> > + return sa->stats.total_time > sb->stats.total_time ? -1 : 1;
> > +}
> > +
> > +int trace_print_bpf_summary(struct syscalltbl *sctbl, FILE *fp)
> > +{
> > + struct syscall_key *prev_key, key;
> > + struct syscall_data *data = NULL;
> > + struct bpf_map *map = skel->maps.syscall_stats_map;
> > + int nr_data = 0;
> > + int printed = 0;
> > +
> > + /* get stats from the bpf map */
> > + prev_key = NULL;
> > + while (!bpf_map__get_next_key(map, prev_key, &key, sizeof(key))) {
> > + struct syscall_stats stat;
> > +
> > + if (!bpf_map__lookup_elem(map, &key, sizeof(key), &stat, sizeof(stat), 0)) {
> > + struct syscall_data *tmp, *pos;
> > +
> > + tmp = realloc(data, sizeof(*data) * (nr_data + 1));
> > + if (tmp == NULL)
> > + break;
> > +
> > + data = tmp;
> > + pos = &data[nr_data++];
> > +
> > + pos->syscall_nr = key.nr;
> > + memcpy(&pos->stats, &stat, sizeof(stat));
> > + }
> > +
> > + prev_key = &key;
> > + }
> > +
> > + qsort(data, nr_data, sizeof(*data), datacmp);
> > +
> > + printed += fprintf(fp, "\n");
> > +
> > + printed += fprintf(fp, " syscall calls errors total min avg max stddev\n");
> > + printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
> > + printed += fprintf(fp, " --------------- -------- ------ -------- --------- --------- --------- ------\n");
> > +
> > + for (int i = 0; i < nr_data; i++) {
> > + struct syscall_data *pos = &data[i];
> > + double total = (double)(pos->stats.total_time) / NSEC_PER_MSEC;
> > + double min = (double)(pos->stats.min_time) / NSEC_PER_MSEC;
> > + double max = (double)(pos->stats.max_time) / NSEC_PER_MSEC;
> > + double avg = total / pos->stats.count;
> > +
> > + printed += fprintf(fp, " %-15s", syscalltbl__name(sctbl, pos->syscall_nr));
> > + printed += fprintf(fp, " %8u %6u %9.3f %9.3f %9.3f %9.3f %9.2f%%\n",
> > + pos->stats.count, pos->stats.error, total, min, avg, max,
> > + /*stddev=*/0.0);
>
> Assuming the performance downfall comes from libc's sqrt(), isn't it
> calculated only after collecting all the data? And don't we have at
> most nr_data records to process? I fail to understand how they "add
> more overheads", can you please explain?
I thought we should do the same with update_stats() in util/stat.c
and do the rest of calculation for stddev in userspace.
void update_stats(struct stats *stats, u64 val)
{
double delta;
stats->n++;
delta = val - stats->mean;
stats->mean += delta / stats->n;
stats->M2 += delta*(val - stats->mean);
if (val > stats->max)
stats->max = val;
if (val < stats->min)
stats->min = val;
}
Then I thought n, mean and M2 should be protected in a lock.
But it may be ok to allow small inaccuracies without locks.
I'll update it.
>
> > + }
> > +
> > + printed += fprintf(fp, "\n\n");
> > + free(data);
> > +
> > + return printed;
> > +}
> > +
> > +void trace_cleanup_bpf_summary(void)
> > +{
> > + syscall_summary_bpf__destroy(skel);
> > +}
> > diff --git a/tools/perf/util/bpf_skel/syscall_summary.bpf.c b/tools/perf/util/bpf_skel/syscall_summary.bpf.c
> > new file mode 100644
> > index 0000000000000000..e573ce39de73eaf3
> > --- /dev/null
> > +++ b/tools/perf/util/bpf_skel/syscall_summary.bpf.c
> > @@ -0,0 +1,109 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * Trace raw_syscalls tracepoints to collect system call statistics.
> > + */
> > +
> > +#include "vmlinux.h"
> > +#include "syscall_summary.h"
> > +
> > +#include <bpf/bpf_helpers.h>
> > +#include <bpf/bpf_tracing.h>
> > +
> > +/* This is to calculate a delta between sys-enter and sys-exit for each thread */
> > +struct syscall_trace {
> > + int nr; /* syscall number is only available at sys-enter */
> > + int unused;
> > + u64 timestamp;
> > +};
> > +
> > +#define MAX_ENTRIES (16 * 1024)
> > +
> > +struct syscall_trace_map {
> > + __uint(type, BPF_MAP_TYPE_HASH);
> > + __type(key, int); /* tid */
> > + __type(value, struct syscall_trace);
> > + __uint(max_entries, MAX_ENTRIES);
> > +} syscall_trace_map SEC(".maps");
> > +
> > +struct syscall_stats_map {
> > + __uint(type, BPF_MAP_TYPE_HASH);
> > + __type(key, struct syscall_key);
> > + __type(value, struct syscall_stats);
> > + __uint(max_entries, MAX_ENTRIES);
> > +} syscall_stats_map SEC(".maps");
> > +
> > +int enabled; /* controlled from userspace */
> > +
> > +static void update_stats(int nr, s64 duration, long ret)
> > +{
> > + struct syscall_key key = { .nr = nr, };
> > + struct syscall_stats *stats;
> > +
> > + stats = bpf_map_lookup_elem(&syscall_stats_map, &key);
> > + if (stats == NULL) {
> > + struct syscall_stats zero = {};
> > +
> > + bpf_map_update_elem(&syscall_stats_map, &key, &zero, BPF_NOEXIST);
> > + stats = bpf_map_lookup_elem(&syscall_stats_map, &key);
> > + if (stats == NULL)
> > + return;
> > + }
> > +
> > + __sync_fetch_and_add(&stats->count, 1);
> > + if (ret < 0)
> > + __sync_fetch_and_add(&stats->error, 1);
> > +
> > + if (duration > 0) {
> > + __sync_fetch_and_add(&stats->total_time, duration);
> > + if (stats->max_time < duration)
> > + stats->max_time = duration;
> > + if (stats->min_time > duration || stats->min_time == 0)
> > + stats->min_time = duration;
> > + }
> > +
> > + return;
> > +}
> > +
> > +SEC("tp_btf/sys_enter")
> > +int sys_enter(u64 *ctx)
> > +{
> > + int tid;
> > + struct syscall_trace st;
> > +
> > + if (!enabled)
> > + return 0;
> > +
> > + st.nr = ctx[1]; /* syscall number */
> > + st.unused = 0;
> > + st.timestamp = bpf_ktime_get_ns();
> > +
> > + tid = bpf_get_current_pid_tgid();
> > + bpf_map_update_elem(&syscall_trace_map, &tid, &st, BPF_ANY);
> > +
> > + return 0;
> > +}
> > +
> > +SEC("tp_btf/sys_exit")
> > +int sys_exit(u64 *ctx)
> > +{
> > + int tid;
> > + long ret = ctx[1]; /* return value of the syscall */
> > + struct syscall_trace *st;
> > + s64 delta;
> > +
> > + if (!enabled)
> > + return 0;
> > +
> > + tid = bpf_get_current_pid_tgid();
> > + st = bpf_map_lookup_elem(&syscall_trace_map, &tid);
> > + if (st == NULL)
> > + return 0;
> > +
> > + delta = bpf_ktime_get_ns() - st->timestamp;
>
> Do you think we can do bpf_ktime_get_ns() a little earlier, say before
> bpf_map_lookup_elem()? Although the difference will be tiny...
Right, also it'd be unnecessary if the lookup returns nothing.
Thanks,
Namhyung
>
>
> > + update_stats(st->nr, delta, ret);
> > +
> > + bpf_map_delete_elem(&syscall_trace_map, &tid);
> > + return 0;
> > +}
> > +
> > +char _license[] SEC("license") = "GPL";
> > diff --git a/tools/perf/util/bpf_skel/syscall_summary.h b/tools/perf/util/bpf_skel/syscall_summary.h
> > new file mode 100644
> > index 0000000000000000..644dc7049377147e
> > --- /dev/null
> > +++ b/tools/perf/util/bpf_skel/syscall_summary.h
> > @@ -0,0 +1,18 @@
> > +// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
> > +/* Data structures shared between BPF and tools. */
> > +#ifndef UTIL_BPF_SKEL_SYSCALL_SUMMARY_H
> > +#define UTIL_BPF_SKEL_SYSCALL_SUMMARY_H
> > +
> > +struct syscall_key {
> > + int nr;
> > +};
> > +
> > +struct syscall_stats {
> > + u64 total_time;
> > + u64 max_time;
> > + u64 min_time;
> > + u32 count;
> > + u32 error;
> > +};
> > +
> > +#endif /* UTIL_BPF_SKEL_SYSCALL_SUMMARY_H */
> > diff --git a/tools/perf/util/trace.h b/tools/perf/util/trace.h
> > new file mode 100644
> > index 0000000000000000..4d7a7c4544d94caf
> > --- /dev/null
> > +++ b/tools/perf/util/trace.h
> > @@ -0,0 +1,31 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +#ifndef UTIL_TRACE_H
> > +#define UTIL_TRACE_H
> > +
> > +#include <stdio.h> /* for FILE */
> > +
> > +struct syscalltbl;
> > +
> > +#ifdef HAVE_BPF_SKEL
> > +
> > +int trace_prepare_bpf_summary(void);
> > +void trace_start_bpf_summary(void);
> > +void trace_end_bpf_summary(void);
> > +int trace_print_bpf_summary(struct syscalltbl *sctbl, FILE *fp);
> > +void trace_cleanup_bpf_summary(void);
> > +
> > +#else /* !HAVE_BPF_SKEL */
> > +
> > +static inline int trace_prepare_bpf_summary(void) { return -1; }
> > +static inline void trace_start_bpf_summary(void) {}
> > +static inline void trace_end_bpf_summary(void) {}
> > +static inline int trace_print_bpf_summary(struct syscalltbl *sctbl __maybe_unused,
> > + FILE *fp __maybe_unused)
> > +{
> > + return 0;
> > +}
> > +static inline void trace_cleanup_bpf_summary(void) {}
> > +
> > +#endif /* HAVE_BPF_SKEL */
> > +
> > +#endif /* UTIL_TRACE_H */
> > --
> > 2.48.1.601.g30ceb7b040-goog
> >
>
> Awesome patch :)
>
> Thanks,
> Howard