perf ftrace: Display latency statistics at the end
Namhyung Kim
maintainerPATCH v2 1/3 perf ftrace: Display latency statistics at the end
Sometimes users also want to see average latency as well as histogram. Display latency statistics like avg, max, min at the end.
$ sudo ./perf ftrace latency -ab -T synchronize_rcu -- ...
# DURATION | COUNT | GRAPH |
0 - 1 us | 0 | |
1 - 2 us | 0 | |
2 - 4 us | 0 | |
4 - 8 us | 0 | |
8 - 16 us | 0 | |
16 - 32 us | 0 | |
32 - 64 us | 0 | |
64 - 128 us | 0 | |
128 - 256 us | 0 | |
256 - 512 us | 0 | |
512 - 1024 us | 0 | |
1 - 2 ms | 0 | |
2 - 4 ms | 0 | |
4 - 8 ms | 0 | |
8 - 16 ms | 1 | ##### |
16 - 32 ms | 7 | ######################################## |
32 - 64 ms | 0 | |
64 - 128 ms | 0 | |
128 - 256 ms | 0 | |
256 - 512 ms | 0 | |
512 - 1024 ms | 0 | |
1 - ... s | 0 | | # statistics (in usec)
total time: 171832
avg time: 21479
max time: 30906
min time: 15869
count: 8--- tools/perf/builtin-ftrace.c | 14 +++++++++++++- tools/perf/util/bpf_ftrace.c | 12 +++++++++++- tools/perf/util/bpf_skel/func_latency.bpf.c | 19 +++++++++++++++++++ tools/perf/util/ftrace.h | 6 ++++-- 4 files changed, 47 insertions(+), 4 deletions(-)
tools/perf/builtin-ftrace.c +13 -1
@@ -43,6 +43,8 @@static volatile sig_atomic_t workload_exec_errno;static volatile sig_atomic_t done; +static struct stats latency_stats;/* for tracepoints */ +static void sig_handler (int sig__maybe_unused ) { done = true;
@@ -798,8 +800,10 @@ static void make_histogram(struct perf_ftrace *ftrace, int buckets[],if (i >=NUM_BUCKET || num >= max_latency - min_latency) i =NUM_BUCKET -1 ; + num += min_latency; do_inc: buckets[i]++; +update_stats (&latency_stats, num); next:/* empty the line buffer for the next output */
@@ -894,6 +898,12 @@ static void display_histogram(struct perf_ftrace *ftrace, int buckets[])printf (" | %10d | %.*s%*s |\n" , buckets[NUM_BUCKET -1 ], bar_len, bar, bar_total - bar_len,"" ); +printf ("\n# statistics (in %s)\n" , ftrace->use_nsec ?"nsec" :"usec" ); +printf (" total time: %20.0f\n" , latency_stats.mean * latency_stats.n); +printf (" avg time: %20.0f\n" , latency_stats.mean); +printf (" max time: %20" PRIu64"\n" , latency_stats.max); +printf (" min time: %20" PRIu64"\n" , latency_stats.min); +printf (" count: %20.0f\n" , latency_stats.n); }static int prepare_func_latency (struct perf_ftrace *ftrace)
@@ -932,6 +942,8 @@ static int prepare_func_latency(struct perf_ftrace *ftrace)if (fd <0 )pr_err ("failed to open trace_pipe\n" ); +init_stats (&latency_stats); +put_tracing_file (trace_file);return fd; }
@@ -961,7 +973,7 @@ static int stop_func_latency(struct perf_ftrace *ftrace)static int read_func_latency (struct perf_ftrace *ftrace,int buckets[]) {if (ftrace->target.use_bpf) -return perf_ftrace__latency_read_bpf (ftrace, buckets); +return perf_ftrace__latency_read_bpf (ftrace, buckets, &latency_stats);return 0 ; }
tools/perf/util/bpf_ftrace.c +11 -1
@@ -11,6 +11,7 @@#include "util/debug.h" #include "util/evlist.h" #include "util/bpf_counter.h" +#include "util/stat.h" #include "util/bpf_skel/func_latency.skel.h"
@@ -86,6 +87,8 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace) } } + skel->bss->min =INT64_MAX ; + skel->links.func_begin =bpf_program__attach_kprobe (skel->progs.func_begin, false, func->name);if (IS_ERR (skel->links.func_begin)) {
@@ -122,7 +125,7 @@ int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused) }int perf_ftrace__latency_read_bpf (struct perf_ftrace *ftrace__maybe_unused , -int buckets[]) +int buckets[],struct stats *stats) {int i, fd, err;u32 idx;
@@ -146,6 +149,13 @@ int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused, buckets[idx] += hist[i]; } +if (skel->bss->count) { + stats->mean = skel->bss->total / skel->bss->count; + stats->n = skel->bss->count; + stats->max = skel->bss->max; + stats->min = skel->bss->min; + } +free (hist);return 0 ; }
tools/perf/util/bpf_skel/func_latency.bpf.c +19 -0
@@ -38,6 +38,12 @@ struct {int enabled =0 ; +// stats +__s64 total; +__s64 count; +__s64 max; +__s64 min; +const volatile int has_cpu =0 ;const volatile int has_task =0 ;const volatile int use_nsec =0 ;
@@ -122,6 +128,8 @@ int BPF_PROG(func_end) delta >= max_latency - min_latency) key =NUM_BUCKET -1 ; } + + delta += min_latency; goto do_lookup; }// calculate index using delta
@@ -136,6 +144,17 @@ int BPF_PROG(func_end)return 0 ; *hist +=1 ; + +if (bucket_range ==0 ) + delta /= cmp_base; + +__sync_fetch_and_add (&total, delta); +__sync_fetch_and_add (&count,1 ); + +if (delta > max) + max = delta; +if (delta < min) + min = delta; }return 0 ;
tools/perf/util/ftrace.h +4 -2
@@ -7,6 +7,7 @@struct evlist ;struct hashamp ; +struct stats ;struct perf_ftrace {struct evlist *evlist;
@@ -46,7 +47,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace);int perf_ftrace__latency_start_bpf (struct perf_ftrace *ftrace);int perf_ftrace__latency_stop_bpf (struct perf_ftrace *ftrace);int perf_ftrace__latency_read_bpf (struct perf_ftrace *ftrace, -int buckets[]); +int buckets[],struct stats *stats);int perf_ftrace__latency_cleanup_bpf (struct perf_ftrace *ftrace);#else /* !HAVE_BPF_SKEL */
@@ -71,7 +72,8 @@ perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused)static inline int perf_ftrace__latency_read_bpf (struct perf_ftrace *ftrace__maybe_unused , -int buckets[]__maybe_unused ) +int buckets[]__maybe_unused , +struct stats *stats__maybe_unused ) {return -1 ; }
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Namhyung Kim
maintainerPATCH v2 2/3 perf ftrace: Add --graph-opts option to profile subcommand
Like trace subcommand, it should be able to pass some options to control the tracing behavior for the function graph tracer. But some options are limited in order to maintain the internal behavior.
For example, it can limit the function call depth like below:
# perf ftrace profile --graph-opts depth=5 -- myprog
--- tools/perf/Documentation/perf-ftrace.txt | 8 ++++++++ tools/perf/builtin-ftrace.c | 4 ++++ 2 files changed, 12 insertions(+)
tools/perf/Documentation/perf-ftrace.txt +8 -0
@@ -201,6 +201,14 @@ OPTIONS for 'perf ftrace profile' Sort the result by the given field. Available values are: total, avg, max, count, name. Default is 'total'. +--graph-opts:: + List of options allowed to set: + + - nosleep-time - Measure on-CPU time only for function_graph tracer. + - noirqs - Ignore functions that happen inside interrupt. + - thresh=<n> - Setup trace duration threshold in microseconds. + - depth=<n> - Set max depth for function graph tracer to follow. + SEE ALSO --------
tools/perf/builtin-ftrace.c +4 -0
@@ -1064,6 +1064,7 @@ static int prepare_func_profile(struct perf_ftrace *ftrace) { ftrace->tracer ="function_graph" ; ftrace->graph_tail =1 ; + ftrace->graph_verbose =0 ; ftrace->profile_hash =hashmap__new (profile_hash, profile_equal,NULL );if (ftrace->profile_hash ==NULL )
@@ -1650,6 +1651,9 @@ int cmd_ftrace(int argc, const char **argv)OPT_CALLBACK ('s' ,"sort" , &profile_sort,"key" ,"Sort result by key: total (default), avg, max, count, name." , parse_sort_key), +OPT_CALLBACK (0 ,"graph-opts" , &ftrace,"options" , +"Graph tracer options, available options: nosleep-time,noirqs,thresh=<n>,depth=<n>" , + parse_graph_tracer_opts), OPT_PARENT(common_options), };const struct option *options = ftrace_options;
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Namhyung Kim
maintainerPATCH v2 3/3 perf test: Update ftrace test to use --graph-opts
In reply to Namhyung KimI found it failed on machines with limited memory because 16M byte per-cpu buffer is too big. The reason it added the option is not to miss tracing data. Thus we can limit the data size by reducing the function call depth instead of increasing the buffer size to handle the whole data.
As it used the same option in the test_ftrace_trace() and it was able to find the sleep function, it should work with the profile subcommand.
Get rid of other grep commands which might be affected by the depth change.
--- tools/perf/tests/shell/ftrace.sh | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-)
tools/perf/tests/shell/ftrace.sh +1 -4
@@ -67,11 +67,8 @@ test_ftrace_latency() { test_ftrace_profile() { echo "perf ftrace profile test" - perf ftrace profile -m 16M sleep 0.1 > "${output}" + perf ftrace profile --graph-opts depth=5 sleep 0.1 > "${output}" grep ^# "${output}" - grep sleep "${output}" - grep schedule "${output}" - grep execve "${output}" time_re="[[:space:]]+1[[:digit:]]{5}\.[[:digit:]]{3}" # 100283.000 100283.000 100283.000 1 __x64_sys_clock_nanosleep # Check for one *clock_nanosleep line with a Count of just 1 that takes a bit more than 0.1 seconds
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Arnaldo Carvalho de Melo
maintainerRe: [PATCH v2 3/3] perf test: Update ftrace test to use --graph-opts
Quoted 6 lines · Namhyung Kim
I found it failed on machines with limited memory because 16M byte per-cpu buffer is too big. The reason it added the option is not to miss tracing data. Thus we can limit the data size by reducing the function call depth instead of increasing the buffer size to handle the whole data.
Quoted 7 lines
As it used the same option in the
test_ftrace_trace()and it was able to find the sleep function, it should work with the profile subcommand.Get rid of other grep commands which might be affected by the depth change.
Cc Thomas Richter <[email protected]>
Hey, I think we should convert the above Cc: to:
As I think you became aware of this situation from a patch proposed by Thomas?
I'm also CCing a few other folks that were CCed in Thomas patch, also CCed here.
- Arnaldo
Quoted 25 lines
Reviewed-by James Clark <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/tests/shell/ftrace.sh | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-)tools/perf/tests/shell/ftrace.sh +1
-4@@ -67,11 +67,8 @@ test_ftrace_latency() { test_ftrace_profile() { echo "perf ftrace profile test" - perf ftrace profile -m 16M sleep 0.1 > "${output}" + perf ftrace profile --graph-opts depth=5 sleep 0.1 > "${output}" grep ^# "${output}" - grep sleep "${output}" - grep schedule "${output}" - grep execve "${output}" time_re="[[:space:]]+1[[:digit:]]{5}\.[[:digit:]]{3}" # 100283.000 100283.000 100283.000 1 __x64_sys_clock_nanosleep # Check for one *clock_nanosleep line with a Count of just 1 that takes a bit more than 0.1 secondsSignature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Quoted 2 lines · Namhyung Kim
Sometimes users also want to see average latency as well as histogram. Display latency statistics like avg, max, min at the end.
Thanks, applied to perf-tools-next, also CCing Gabriele, since he did work on this tool recently.
- Arnaldo
Quoted 217 lines
$ sudo ./perf ftrace latency -ab -T synchronize_rcu -- ... # DURATION | COUNT | GRAPH | 0 - 1 us | 0 | | 1 - 2 us | 0 | | 2 - 4 us | 0 | | 4 - 8 us | 0 | | 8 - 16 us | 0 | | 16 - 32 us | 0 | | 32 - 64 us | 0 | | 64 - 128 us | 0 | | 128 - 256 us | 0 | | 256 - 512 us | 0 | | 512 - 1024 us | 0 | | 1 - 2 ms | 0 | | 2 - 4 ms | 0 | | 4 - 8 ms | 0 | | 8 - 16 ms | 1 | ##### | 16 - 32 ms | 7 | ######################################## | 32 - 64 ms | 0 | | 64 - 128 ms | 0 | | 128 - 256 ms | 0 | | 256 - 512 ms | 0 | | 512 - 1024 ms | 0 | | 1 - ... s | 0 | |# statistics (in usec) total time: 171832 avg time: 21479 max time: 30906 min time: 15869 count: 8Reviewed-by James Clark <[email protected]>Signed-off-by Namhyung Kim <[email protected]>--- tools/perf/builtin-ftrace.c | 14 +++++++++++++- tools/perf/util/bpf_ftrace.c | 12 +++++++++++- tools/perf/util/bpf_skel/func_latency.bpf.c | 19 +++++++++++++++++++ tools/perf/util/ftrace.h | 6 ++++-- 4 files changed, 47 insertions(+), 4 deletions(-)tools/perf/builtin-ftrace.c +13
-1@@ -43,6 +43,8 @@static volatile sig_atomic_t workload_exec_errno;static volatile sig_atomic_t done; +static struct stats latency_stats;/* for tracepoints */ +static void sig_handler (int sig__maybe_unused ) { done = true;@@ -798,8 +800,10 @@ static void make_histogram(struct perf_ftrace *ftrace, int buckets[],if (i >=NUM_BUCKET || num >= max_latency - min_latency) i =NUM_BUCKET -1 ; + num += min_latency; do_inc: buckets[i]++; +update_stats (&latency_stats, num); next:/* empty the line buffer for the next output */ @@ -894,6 +898,12 @@ static void display_histogram(struct perf_ftrace *ftrace, int buckets[])printf (" | %10d | %.*s%*s |\n" , buckets[NUM_BUCKET -1 ], bar_len, bar, bar_total - bar_len,"" ); +printf ("\n# statistics (in %s)\n" , ftrace->use_nsec ?"nsec" :"usec" ); +printf (" total time: %20.0f\n" , latency_stats.mean * latency_stats.n); +printf (" avg time: %20.0f\n" , latency_stats.mean); +printf (" max time: %20" PRIu64"\n" , latency_stats.max); +printf (" min time: %20" PRIu64"\n" , latency_stats.min); +printf (" count: %20.0f\n" , latency_stats.n); }static int prepare_func_latency (struct perf_ftrace *ftrace)@@ -932,6 +942,8 @@ static int prepare_func_latency(struct perf_ftrace *ftrace)if (fd <0 )pr_err ("failed to open trace_pipe\n" ); +init_stats (&latency_stats); +put_tracing_file (trace_file);return fd; }@@ -961,7 +973,7 @@ static int stop_func_latency(struct perf_ftrace *ftrace)static int read_func_latency (struct perf_ftrace *ftrace,int buckets[]) {if (ftrace->target.use_bpf) -return perf_ftrace__latency_read_bpf (ftrace, buckets); +return perf_ftrace__latency_read_bpf (ftrace, buckets, &latency_stats);return 0 ; }tools/perf/util/bpf_ftrace.c +11
-1@@ -11,6 +11,7 @@#include "util/debug.h" #include "util/evlist.h" #include "util/bpf_counter.h" +#include "util/stat.h" #include "util/bpf_skel/func_latency.skel.h" @@ -86,6 +87,8 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace) } } + skel->bss->min =INT64_MAX ; + skel->links.func_begin =bpf_program__attach_kprobe (skel->progs.func_begin, false, func->name);if (IS_ERR (skel->links.func_begin)) {@@ -122,7 +125,7 @@ int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused) }int perf_ftrace__latency_read_bpf (struct perf_ftrace *ftrace__maybe_unused , -int buckets[]) +int buckets[],struct stats *stats) {int i, fd, err;u32 idx;@@ -146,6 +149,13 @@ int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused, buckets[idx] += hist[i]; } +if (skel->bss->count) { + stats->mean = skel->bss->total / skel->bss->count; + stats->n = skel->bss->count; + stats->max = skel->bss->max; + stats->min = skel->bss->min; + } +free (hist);return 0 ; }tools/perf/util/bpf_skel/func_latency.bpf.c +19
-0@@ -38,6 +38,12 @@ struct {int enabled =0 ; +// stats +__s64 total; +__s64 count; +__s64 max; +__s64 min; +const volatile int has_cpu =0 ;const volatile int has_task =0 ;const volatile int use_nsec =0 ;@@ -122,6 +128,8 @@ int BPF_PROG(func_end) delta >= max_latency - min_latency) key =NUM_BUCKET -1 ; } + + delta += min_latency; goto do_lookup; }// calculate index using delta @@ -136,6 +144,17 @@ int BPF_PROG(func_end)return 0 ; *hist +=1 ; + +if (bucket_range ==0 ) + delta /= cmp_base; + +__sync_fetch_and_add (&total, delta); +__sync_fetch_and_add (&count,1 ); + +if (delta > max) + max = delta; +if (delta < min) + min = delta; }return 0 ;tools/perf/util/ftrace.h +4
-2@@ -7,6 +7,7 @@struct evlist ;struct hashamp ; +struct stats ;struct perf_ftrace {struct evlist *evlist;@@ -46,7 +47,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace);int perf_ftrace__latency_start_bpf (struct perf_ftrace *ftrace);int perf_ftrace__latency_stop_bpf (struct perf_ftrace *ftrace);int perf_ftrace__latency_read_bpf (struct perf_ftrace *ftrace, -int buckets[]); +int buckets[],struct stats *stats);int perf_ftrace__latency_cleanup_bpf (struct perf_ftrace *ftrace);#else /* !HAVE_BPF_SKEL */ @@ -71,7 +72,8 @@ perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused)static inline int perf_ftrace__latency_read_bpf (struct perf_ftrace *ftrace__maybe_unused , -int buckets[]__maybe_unused ) +int buckets[]__maybe_unused , +struct stats *stats__maybe_unused ) {return -1 ; }Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog