perf report/annotate: Add option to specify a CPU range
PATCH 1/1 perf report/annotate: Add option to specify a CPU range
Add an option to perf report and perf annotate to specify which CPUs to operate on. This enables us to take a single system wide profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the bottleneck was on one CPU but this was hidden in the overall profile. Per process and per thread breakdowns didn't help because multiple processes were running on each CPU and no single process consumed an entire CPU.
The patch converts the list of CPUs returned by cpu_map__new into a bitmap for fast lookup. I wanted to use -C to be consistent with perf top/record/stat, but unfortunately perf report already uses -C <comms>.
---
I capped it at MAX_NR_CPUS to avoid having to dynamically allocate cpu_bitmap, but we could do that if the extra complexity is worth it.
tools/perf/builtin-report.c +27 -0
@@ -33,6 +33,9 @@#include "util/sort.h" #include "util/hist.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -48,6 +51,9 @@ static const char *pretty_printing_stylestatic char callchain_default_opt[] ="fractal,0.5" ;static symbol_filter_t annotate_init; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_session__add_hist_entry(struct perf_session *session, struct addr_location *al, struct perf_sample *sample,
@@ -116,6 +122,9 @@ static int process_sample_event(union peif (al.filtered || (hide_unresolved && al.sym ==NULL ))return 0 ; +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (al.map !=NULL ) al.map->dso->hit =1 ;
@@ -455,6 +464,7 @@ static const struct option options[] = { "Only display entries resolved to a symbol"),OPT_STRING (0 ,"symfs" , &symbol_conf.symfs,"directory" ,"Look for files with symbols relative to this directory" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
@@ -501,6 +511,23 @@ int cmd_report(int argc, const char **arsetup_sorting (report_usage, options); +if (cpu_list) { +int i; +struct cpu_map *map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +fprintf (stderr,"Requested CPU %d too large, " +"consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + } +if (parent_pattern != default_parent_pattern) {if (sort_dimension__add ("parent" ) <0 )return -1 ;
tools/perf/builtin-annotate.c +27 -0
@@ -28,6 +28,9 @@#include "util/hist.h" #include "util/session.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -38,6 +41,9 @@ static bool print_line;static const char *sym_hist_filter; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_evlist__add_sample(struct perf_evlist *evlist, struct perf_sample *sample, struct perf_evsel *evsel,
@@ -90,6 +96,9 @@ static int process_sample_event(union pereturn -1 ; } +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (!al.filtered &&perf_evlist__add_sample (session->evlist, sample, evsel, &al)) { pr_warning("problem incrementing symbol count, "
@@ -252,6 +261,7 @@ static const struct option options[] = { "print matching source lines (may be slow)"),OPT_BOOLEAN ('P' ,"full-paths" , &full_paths,"Don't shorten the displayed pathnames" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
@@ -274,6 +284,23 @@ int cmd_annotate(int argc, const char **setup_sorting (annotate_usage, options); +if (cpu_list) { +int i; +struct cpu_map *map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +fprintf (stderr,"Requested CPU %d too large, " +"consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + } +if (argc) {/* * Special case: if there's an argument left then assume tha
Quoted 2 lines · Anton Blanchard
Add an option to perf report and perf annotate to specify which CPUs
What about perf-script?
Quoted 44 lines
to operate on. This enables us to take a single system wide profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the bottleneck was on one CPU but this was hidden in the overall profile. Per process and per thread breakdowns didn't help because multiple processes were running on each CPU and no single process consumed an entire CPU.
The patch converts the list of CPUs returned by
cpu_map__newinto a bitmap for fast lookup. I wanted to use -C to be consistent with perf top/record/stat, but unfortunately perf report already uses -C <comms>.Signed-off-by Anton Blanchard <[email protected]>---
I capped it at
MAX_NR_CPUSto avoid having to dynamically allocatecpu_bitmap, but we could do that if the extra complexity is worth it.Index:
linux-2.6-tip/tools/perf/builtin-report.c
===================================================================tools/perf/builtin-report.c +6
-0@@ -33,6 +33,9 @@#include "util/sort.h" #include "util/hist.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;@@ -48,6 +51,9 @@ static const char *pretty_printing_stylestatic char callchain_default_opt[] ="fractal,0.5" ;static symbol_filter_t annotate_init; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_session__add_hist_entry(struct perf_session *session, struct addr_location *al, struct perf_sample *sample,
tools/perf/builtin-report.c static int process_sample_event(union pe · line 127
Quoted 6 lines
tools/perf/builtin-report.c +3
-0@@ -116,6 +122,9 @@ static int process_sample_event(union peif (al.filtered || (hide_unresolved && al.sym ==NULL ))return 0 ; +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +
Need to check that the SAMPLE_CPU attribute is set for the event for which the sample is generated; see builtin-script.c, perf_evsel__check_attr().
tools/perf/builtin-report.c static const struct option options[] = { · line 467
Quoted 8 lines
tools/perf/builtin-report.c +1
-0if (al.map !=NULL ) al.map->dso->hit =1 ;@@ -455,6 +464,7 @@ static const struct option options[] = { "Only display entries resolved to a symbol"),OPT_STRING (0 ,"symfs" , &symbol_conf.symfs,"directory" ,"Look for files with symbols relative to this directory" ), + OPT_STRING('c', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
This option should be added to the Documentation file, Documentation/perf-report.txt
Same comments for perf-annotate changes below.
David
Quoted 98 lines
tools/perf/builtin-report.c +17
-0OPT_END() };@@ -501,6 +511,23 @@ int cmd_report(int argc, const char **arsetup_sorting (report_usage, options); +if (cpu_list) { +int i; +struct cpu_map *map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +fprintf (stderr,"Requested CPU %d too large, " +"consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + } +if (parent_pattern != default_parent_pattern) {if (sort_dimension__add ("parent" ) <0 )return -1 ;Index:
linux-2.6-tip/tools/perf/builtin-annotate.c
===================================================================tools/perf/builtin-annotate.c +27
-1@@ -28,6 +28,9 @@#include "util/hist.h" #include "util/session.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;@@ -38,6 +41,9 @@ static bool print_line;static const char *sym_hist_filter; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_evlist__add_sample(struct perf_evlist *evlist, struct perf_sample *sample, struct perf_evsel *evsel,@@ -90,6 +96,9 @@ static int process_sample_event(union pereturn -1 ; } +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (!al.filtered &&perf_evlist__add_sample (session->evlist, sample, evsel, &al)) { pr_warning("problem incrementing symbol count, "@@ -252,6 +261,7 @@ static const struct option options[] = { "print matching source lines (may be slow)"),OPT_BOOLEAN ('P' ,"full-paths" , &full_paths,"Don't shorten the displayed pathnames" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };@@ -274,6 +284,23 @@ int cmd_annotate(int argc, const char **setup_sorting (annotate_usage, options); +if (cpu_list) { +int i; +struct cpu_map *map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +fprintf (stderr,"Requested CPU %d too large, " +"consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + } +if (argc) {/* * Special case: if there's an argument left then assume tha -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Hi David,
Quoted 3 lines
Quoted 1 line
Add an option to perf report and perf annotate to specify which CPUs
What about perf-script?
Good point, I added the option to perf script.
Quoted 7 lines
tools/perf/builtin-report.c static int process_sample_event(union pe · line 127
Quoted 3 lines
tools/perf/builtin-report.c +3
-0+if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +Need to check that the
SAMPLE_CPUattribute is set for the event for which the sample is generated; seebuiltin-script.c,perf_evsel__check_attr().
Added.
Quoted 2 lines
This option should be added to the Documentation file,
Documentation/perf-report.txt
Added. Updated patch on the way.
Thanks,
Anton
PATCH 1/1 perf report/annotate: Add option to specify a CPU range
In reply to David AhernAdd an option to perf report/annotate/script to specify which CPUs to operate on. This enables us to take a single system wide profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the bottleneck was on one CPU but this was hidden in the overall profile. Per process and per thread breakdowns didn't help because multiple processes were running on each CPU and no single process consumed an entire CPU.
The patch converts the list of CPUs returned by cpu_map__new into a bitmap for fast lookup. I wanted to use -C to be consistent with perf top/record/stat, but unfortunately perf report already uses -C <comms>.
---
v2: Incorporate suggestions from David Ahern:
- Added -c to perf script
- Check that SAMPLE_CPU is set when -c is used
- Update documentation
tools/perf/builtin-report.c +45 -0
@@ -33,6 +33,9 @@#include "util/sort.h" #include "util/hist.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -48,6 +51,9 @@ static const char *pretty_printing_stylestatic char callchain_default_opt[] ="fractal,0.5" ;static symbol_filter_t annotate_init; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_session__add_hist_entry(struct perf_session *session, struct addr_location *al, struct perf_sample *sample,
@@ -116,6 +122,9 @@ static int process_sample_event(union peif (al.filtered || (hide_unresolved && al.sym ==NULL ))return 0 ; +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (al.map !=NULL ) al.map->dso->hit =1 ;
@@ -262,6 +271,41 @@ static int __cmd_report(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { +int i; +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); + ret =-1 ; + goto out_delete; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); + ret =-1 ; + goto out_delete; + } + +set_bit (cpu, cpu_bitmap); + } + } +if (show_threads)perf_read_values_init (&show_threads_values);
@@ -455,6 +499,7 @@ static const struct option options[] = { "Only display entries resolved to a symbol"),OPT_STRING (0 ,"symfs" , &symbol_conf.symfs,"directory" ,"Look for files with symbols relative to this directory" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
tools/perf/builtin-annotate.c +45 -0
75 lines not on this page · raw message
tools/perf/builtin-script.c +41 -0
71 lines not on this page · raw message
tools/perf/Documentation/perf-annotate.txt +6 -0
@@ -66,6 +66,12 @@ OPTIONS used. This interfaces starts by centering on the line with more samples, TAB/UNTAB cycles through the lines with more samples. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-report[1]
tools/perf/Documentation/perf-report.txt +6 -0
@@ -119,6 +119,12 @@ OPTIONS --symfs=<directory>:: Look for files with symbols relative to this directory. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-stat[1]
tools/perf/Documentation/perf-script.txt +6 -0
@@ -182,6 +182,12 @@ OPTIONS --hide-call-graph:: When printing symbols do not display call chain. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-script-perl[1],
Quoted 86 lines · Anton Blanchard
Add an option to perf report/annotate/script to specify which CPUs to operate on. This enables us to take a single system wide profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the bottleneck was on one CPU but this was hidden in the overall profile. Per process and per thread breakdowns didn't help because multiple processes were running on each CPU and no single process consumed an entire CPU.
The patch converts the list of CPUs returned by
cpu_map__newinto a bitmap for fast lookup. I wanted to use -C to be consistent with perf top/record/stat, but unfortunately perf report already uses -C <comms>.Signed-off-by Anton Blanchard <[email protected]>---
v2: Incorporate suggestions from David Ahern:
- Added -c to perf script
- Check thatSAMPLE_CPUis set when -c is used
- Update documentationIndex:
linux-2.6-tip/tools/perf/builtin-report.c
===================================================================tools/perf/builtin-report.c +34
-0@@ -33,6 +33,9 @@#include "util/sort.h" #include "util/hist.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;@@ -48,6 +51,9 @@ static const char *pretty_printing_stylestatic char callchain_default_opt[] ="fractal,0.5" ;static symbol_filter_t annotate_init; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_session__add_hist_entry(struct perf_session *session, struct addr_location *al, struct perf_sample *sample,@@ -116,6 +122,9 @@ static int process_sample_event(union peif (al.filtered || (hide_unresolved && al.sym ==NULL ))return 0 ; +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (al.map !=NULL ) al.map->dso->hit =1 ;@@ -262,6 +271,41 @@ static int __cmd_report(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { +int i; +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); + ret =-1 ; + goto out_delete; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) {
tools/perf/builtin-report.c static int __cmd_report(void) · line 308
Quoted 10 lines
tools/perf/builtin-report.c +10
-0+pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); + ret =-1 ; + goto out_delete; + } + +set_bit (cpu, cpu_bitmap); + } + } +
It would be better to make this a function that all 3 commands reference -- something like perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap) in util/session.c
David
Quoted 227 lines
tools/perf/builtin-report.c +1
-0if (show_threads)perf_read_values_init (&show_threads_values);@@ -455,6 +499,7 @@ static const struct option options[] = { "Only display entries resolved to a symbol"),OPT_STRING (0 ,"symfs" , &symbol_conf.symfs,"directory" ,"Look for files with symbols relative to this directory" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };Index:
linux-2.6-tip/tools/perf/builtin-annotate.c
===================================================================tools/perf/builtin-annotate.c +45
-0@@ -28,6 +28,9 @@#include "util/hist.h" #include "util/session.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;@@ -38,6 +41,9 @@ static bool print_line;static const char *sym_hist_filter; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_evlist__add_sample(struct perf_evlist *evlist, struct perf_sample *sample, struct perf_evsel *evsel,@@ -90,6 +96,9 @@ static int process_sample_event(union pereturn -1 ; } +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (!al.filtered &&perf_evlist__add_sample (session->evlist, sample, evsel, &al)) { pr_warning("problem incrementing symbol count, "@@ -177,6 +186,41 @@ static int __cmd_annotate(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { +int i; +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); + ret =-1 ; + goto out_delete; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); + ret =-1 ; + goto out_delete; + } + +set_bit (cpu, cpu_bitmap); + } + } + ret =perf_session__process_events (session, &event_ops);if (ret) goto out_delete;@@ -252,6 +296,7 @@ static const struct option options[] = { "print matching source lines (may be slow)"),OPT_BOOLEAN ('P' ,"full-paths" , &full_paths,"Don't shorten the displayed pathnames" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };Index:
linux-2.6-tip/tools/perf/builtin-script.c
===================================================================tools/perf/builtin-script.c +41
-0@@ -13,6 +13,8 @@#include "util/util.h" #include "util/evlist.h" #include "util/evsel.h" +#include <linux/bitmap.h> +#include "util/cpumap.h" static char const *script_name;static char const *generate_script_lang;@@ -21,6 +23,8 @@ static u64 last_timestamp;static u64 nr_unordered;extern const struct option record_options[];static bool no_callchain; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);enum perf_output_field {PERF_OUTPUT_COMM =1U <<0 ,@@ -453,6 +457,10 @@ static int process_sample_event(union pe last_timestamp = sample->time;return 0 ; } + +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; + scripting_ops->process_event (event, sample, evsel, session, thread); session->hists.stats.total_period += sample->period;@@ -1075,6 +1083,7 @@ static const struct option options[] = {OPT_CALLBACK ('f' ,"fields" ,NULL ,"str" ,"comma separated output fields prepend with 'type:'. Valid types: hw,sw,trace,raw. Fields: comm,tid,pid,time,cpu,event,trace,ip,sym,dso,addr" , parse_output_fields), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };@@ -1255,6 +1264,38 @@ int cmd_script(int argc, const char **arif (session ==NULL )return -ENOMEM ; +if (cpu_list) { +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); +return -1 ; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + } +if (!no_callchain) symbol_conf.use_callchain = true;else Index:
linux-2.6-tip/tools/perf/Documentation/perf-annotate.txt
===================================================================tools/perf/Documentation/perf-annotate.txt +6
-0@@ -66,6 +66,12 @@ OPTIONS used. This interfaces starts by centering on the line with more samples, TAB/UNTAB cycles through the lines with more samples. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-report[1]Index:
linux-2.6-tip/tools/perf/Documentation/perf-report.txt
===================================================================tools/perf/Documentation/perf-report.txt +6
-0@@ -119,6 +119,12 @@ OPTIONS --symfs=<directory>:: Look for files with symbols relative to this directory. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-stat[1]Index:
linux-2.6-tip/tools/perf/Documentation/perf-script.txt
===================================================================tools/perf/Documentation/perf-script.txt +6
-0@@ -182,6 +182,12 @@ OPTIONS --hide-call-graph:: When printing symbols do not display call chain. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-script-perl[1],
Quoted 35 lines · David Ahern
Quoted 35 lines
tools/perf/builtin-report.c +31
-0@@ -262,6 +271,41 @@ static int __cmd_report(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { +int i; +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); + ret =-1 ; + goto out_delete; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); + ret =-1 ; + goto out_delete; + } +
Quoted 8 lines
tools/perf/builtin-report.c static int __cmd_report(void) · line 308
Quoted 4 lines
tools/perf/builtin-report.c +4
-0+set_bit (cpu, cpu_bitmap); + } + } +It would be better to make this a function that all 3 commands reference -- something like
perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap)inutil/session.c
Agreed. I can see how it ended up looking like this (fixing only perf report, then adding it to top, then to script), but at this stage it really calls for one helper that all three commands can utilize.
Very nice enhancement otherwise - might i suggest a 'perf top' hotkey as well to limit the output to certain CPUs only? :-)
Thanks,
Ingo
Hi,
Quoted 8 lines · Ingo Molnar
Quoted 3 lines · David Ahern
It would be better to make this a function that all 3 commands reference -- something like
perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap)inutil/session.cAgreed. I can see how it ended up looking like this (fixing only perf report, then adding it to top, then to script), but at this stage it really calls for one helper that all three commands can utilize.
Yeah, it was dying to be consolidated. New version on the way.
Quoted 2 lines
Very nice enhancement otherwise - might i suggest a 'perf top' hotkey as well to limit the output to certain CPUs only? :-)
Good idea :) I'll have a go at it.
Anton
PATCH 1/1 perf report/annotate/script: Add option to specify a CPU range
In reply to Ingo MolnarAdd an option to perf report/annotate/script to specify which CPUs to operate on. This enables us to take a single system wide profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the bottleneck was on one CPU but this was hidden in the overall profile. Per process and per thread breakdowns didn't help because multiple processes were running on each CPU and no single process consumed an entire CPU.
The patch converts the list of CPUs returned by cpu_map__new into a bitmap for fast lookup. I wanted to use -C to be consistent with perf top/record/stat, but unfortunately perf report already uses -C <comms>.
---
v2: Incorporate suggestions from David Ahern:
- Added -c to perf script
- Check that SAMPLE_CPU is set when -c is used
- Update documentation
v3: Create perf_session__cpu_bitmap
tools/perf/builtin-report.c +15 -0
@@ -33,6 +33,8 @@#include "util/sort.h" #include "util/hist.h" +#include <linux/bitmap.h> +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -49,6 +51,9 @@ static char callchain_default_opt[] = "static bool inverted_callchain;static symbol_filter_t annotate_init; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_session__add_hist_entry(struct perf_session *session, struct addr_location *al, struct perf_sample *sample,
@@ -117,6 +122,9 @@ static int process_sample_event(union peif (al.filtered || (hide_unresolved && al.sym ==NULL ))return 0 ; +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (al.map !=NULL ) al.map->dso->hit =1 ;
@@ -263,6 +271,12 @@ static int __cmd_report(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { + ret =perf_session__cpu_bitmap (session, cpu_list, cpu_bitmap); +if (ret) + goto out_delete; + } +if (show_threads)perf_read_values_init (&show_threads_values);
@@ -473,6 +487,7 @@ static const struct option options[] = { "Only display entries resolved to a symbol"),OPT_STRING (0 ,"symfs" , &symbol_conf.symfs,"directory" ,"Look for files with symbols relative to this directory" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
tools/perf/builtin-annotate.c +15 -0
@@ -28,6 +28,8 @@#include "util/hist.h" #include "util/session.h" +#include <linux/bitmap.h> +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -38,6 +40,9 @@ static bool print_line;static const char *sym_hist_filter; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_evlist__add_sample(struct perf_evlist *evlist, struct perf_sample *sample, struct perf_evsel *evsel,
@@ -90,6 +95,9 @@ static int process_sample_event(union pereturn -1 ; } +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (!al.filtered &&perf_evlist__add_sample (session->evlist, sample, evsel, &al)) { pr_warning("problem incrementing symbol count, "
@@ -177,6 +185,12 @@ static int __cmd_annotate(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { + ret =perf_session__cpu_bitmap (session, cpu_list, cpu_bitmap); +if (ret) + goto out_delete; + } + ret =perf_session__process_events (session, &event_ops);if (ret) goto out_delete;
@@ -252,6 +266,7 @@ static const struct option options[] = { "print matching source lines (may be slow)"),OPT_BOOLEAN ('P' ,"full-paths" , &full_paths,"Don't shorten the displayed pathnames" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
tools/perf/builtin-script.c +13 -0
@@ -13,6 +13,7 @@#include "util/util.h" #include "util/evlist.h" #include "util/evsel.h" +#include <linux/bitmap.h> static char const *script_name;static char const *generate_script_lang;
@@ -21,6 +22,8 @@ static u64 last_timestamp;static u64 nr_unordered;extern const struct option record_options[];static bool no_callchain; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);enum perf_output_field {PERF_OUTPUT_COMM =1U <<0 ,
@@ -453,6 +456,10 @@ static int process_sample_event(union pe last_timestamp = sample->time;return 0 ; } + +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; + scripting_ops->process_event (event, sample, evsel, session, thread); session->hists.stats.total_period += sample->period;
@@ -1075,6 +1082,7 @@ static const struct option options[] = {OPT_CALLBACK ('f' ,"fields" ,NULL ,"str" ,"comma separated output fields prepend with 'type:'. Valid types: hw,sw,trace,raw. Fields: comm,tid,pid,time,cpu,event,trace,ip,sym,dso,addr" , parse_output_fields), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
@@ -1255,6 +1263,11 @@ int cmd_script(int argc, const char **arif (session ==NULL )return -ENOMEM ; +if (cpu_list) { +if (perf_session__cpu_bitmap (session, cpu_list, cpu_bitmap)) +return -1 ; + } +if (!no_callchain) symbol_conf.use_callchain = true;else
tools/perf/Documentation/perf-annotate.txt +6 -0
@@ -66,6 +66,12 @@ OPTIONS used. This interfaces starts by centering on the line with more samples, TAB/UNTAB cycles through the lines with more samples. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-report[1]
tools/perf/Documentation/perf-report.txt +6 -0
@@ -128,6 +128,12 @@ OPTIONS --symfs=<directory>:: Look for files with symbols relative to this directory. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-stat[1]
tools/perf/Documentation/perf-script.txt +6 -0
@@ -182,6 +182,12 @@ OPTIONS --hide-call-graph:: When printing symbols do not display call chain. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-script-perl[1],
tools/perf/util/session.c +38 -0
@@ -12,6 +12,7 @@#include "session.h" #include "sort.h" #include "util.h" +#include "cpumap.h" static int perf_session__open (struct perf_session *self,bool force) {
@@ -1282,3 +1283,40 @@ void perf_session__print_ip(union perf_e } } } + +int perf_session__cpu_bitmap (struct perf_session *session, +const char *cpu_list,unsigned long *cpu_bitmap) +{ +int i; +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); +return -1 ; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + +return 0 ; +}
tools/perf/util/session.h +3 -0
@@ -172,4 +172,7 @@ void perf_session__print_ip(union perf_estruct perf_session *session, int print_sym, int print_dso); +int perf_session__cpu_bitmap (struct perf_session *session, +const char *cpu_list,unsigned long *cpu_bitmap); +#endif /* __PERF_SESSION_H */
[tip:perf/core] perf report/annotate/script: Add option to specify a CPU range
Commit-ID: 5d67be97f8903d05ce53597fb5f3bc25a45e8026 Gitweb: http://git.kernel.org/tip/5d67be97f8903d05ce53597fb5f3bc25a45e8026 Author: Anton Blanchard <[email protected]> AuthorDate: Mon, 4 Jul 2011 21:57:50 +1000 Committer: Ingo Molnar <[email protected]> CommitDate: Tue, 5 Jul 2011 10:44:44 +0200
perf report/annotate/script: Add option to specify a CPU range
Add an option to perf report/annotate/script to specify which CPUs to operate on. This enables us to take a single system wide profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the bottleneck was on one CPU but this was hidden in the overall profile. Per process and per thread breakdowns didn't help because multiple processes were running on each CPU and no single process consumed an entire CPU.
The patch converts the list of CPUs returned by cpu_map__new into a bitmap for fast lookup. I wanted to use -C to be consistent with perf top/record/stat, but unfortunately perf report already uses -C <comms>.
v2: Incorporate suggestions from David Ahern:
- Added -c to perf script
- Check that SAMPLE_CPU is set when -c is used
- Update documentation
v3: Create perf_session__cpu_bitmap()
--- tools/perf/Documentation/perf-annotate.txt | 6 ++++ tools/perf/Documentation/perf-report.txt | 6 ++++ tools/perf/Documentation/perf-script.txt | 6 ++++ tools/perf/builtin-annotate.c | 15 +++++++++++ tools/perf/builtin-report.c | 15 +++++++++++ tools/perf/builtin-script.c | 13 +++++++++ tools/perf/util/session.c | 38 ++++++++++++++++++++++++++++ tools/perf/util/session.h | 3 ++ 8 files changed, 102 insertions(+), 0 deletions(-)
tools/perf/Documentation/perf-annotate.txt +6 -0
@@ -66,6 +66,12 @@ OPTIONS used. This interfaces starts by centering on the line with more samples, TAB/UNTAB cycles through the lines with more samples. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-report[1]
tools/perf/Documentation/perf-report.txt +6 -0
@@ -128,6 +128,12 @@ OPTIONS --symfs=<directory>:: Look for files with symbols relative to this directory. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-stat[1]
tools/perf/Documentation/perf-script.txt +6 -0
@@ -182,6 +182,12 @@ OPTIONS --hide-call-graph:: When printing symbols do not display call chain. +-c:: +--cpu:: Only report samples for the list of CPUs provided. Multiple CPUs can + be provided as a comma-separated list with no space: 0,1. Ranges of + CPUs are specified with -: 0-2. Default is to report samples on all + CPUs. + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-script-perl[1],
tools/perf/builtin-annotate.c +15 -0
@@ -28,6 +28,8 @@#include "util/hist.h" #include "util/session.h" +#include <linux/bitmap.h> +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -38,6 +40,9 @@ static bool print_line;static const char *sym_hist_filter; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_evlist__add_sample(struct perf_evlist *evlist, struct perf_sample *sample, struct perf_evsel *evsel,
@@ -90,6 +95,9 @@ static int process_sample_event(union perf_event *event,return -1 ; } +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (!al.filtered &&perf_evlist__add_sample (session->evlist, sample, evsel, &al)) { pr_warning("problem incrementing symbol count, "
@@ -177,6 +185,12 @@ static int __cmd_annotate(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { + ret =perf_session__cpu_bitmap (session, cpu_list, cpu_bitmap); +if (ret) + goto out_delete; + } + ret =perf_session__process_events (session, &event_ops);if (ret) goto out_delete;
@@ -252,6 +266,7 @@ static const struct option options[] = { "print matching source lines (may be slow)"),OPT_BOOLEAN ('P' ,"full-paths" , &full_paths,"Don't shorten the displayed pathnames" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
tools/perf/builtin-report.c +15 -0
@@ -33,6 +33,8 @@#include "util/sort.h" #include "util/hist.h" +#include <linux/bitmap.h> +static char const *input_name ="perf.data" ;static bool force, use_tui, use_stdio;
@@ -49,6 +51,9 @@ static char callchain_default_opt[] = "fractal,0.5,callee";static bool inverted_callchain;static symbol_filter_t annotate_init; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); + static int perf_session__add_hist_entry(struct perf_session *session, struct addr_location *al, struct perf_sample *sample,
@@ -117,6 +122,9 @@ static int process_sample_event(union perf_event *event,if (al.filtered || (hide_unresolved && al.sym ==NULL ))return 0 ; +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; +if (al.map !=NULL ) al.map->dso->hit =1 ;
@@ -263,6 +271,12 @@ static int __cmd_report(void)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { + ret =perf_session__cpu_bitmap (session, cpu_list, cpu_bitmap); +if (ret) + goto out_delete; + } +if (show_threads)perf_read_values_init (&show_threads_values);
@@ -473,6 +487,7 @@ static const struct option options[] = { "Only display entries resolved to a symbol"),OPT_STRING (0 ,"symfs" , &symbol_conf.symfs,"directory" ,"Look for files with symbols relative to this directory" ), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
tools/perf/builtin-script.c +13 -0
@@ -13,6 +13,7 @@#include "util/util.h" #include "util/evlist.h" #include "util/evsel.h" +#include <linux/bitmap.h> static char const *script_name;static char const *generate_script_lang;
@@ -21,6 +22,8 @@ static u64 last_timestamp;static u64 nr_unordered;extern const struct option record_options[];static bool no_callchain; +static const char *cpu_list; +static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);enum perf_output_field {PERF_OUTPUT_COMM =1U <<0 ,
@@ -453,6 +456,10 @@ static int process_sample_event(union perf_event *event, last_timestamp = sample->time;return 0 ; } + +if (cpu_list && !test_bit (sample->cpu, cpu_bitmap)) +return 0 ; + scripting_ops->process_event (event, sample, evsel, session, thread); session->hists.stats.total_period += sample->period;
@@ -1075,6 +1082,7 @@ static const struct option options[] = {OPT_CALLBACK ('f' ,"fields" ,NULL ,"str" ,"comma separated output fields prepend with 'type:'. Valid types: hw,sw,trace,raw. Fields: comm,tid,pid,time,cpu,event,trace,ip,sym,dso,addr" , parse_output_fields), +OPT_STRING ('c' ,"cpu" , &cpu_list,"cpu" ,"list of cpus to profile" ), OPT_END() };
@@ -1255,6 +1263,11 @@ int cmd_script(int argc, const char **argv, const char *prefix __used)if (session ==NULL )return -ENOMEM ; +if (cpu_list) { +if (perf_session__cpu_bitmap (session, cpu_list, cpu_bitmap)) +return -1 ; + } +if (!no_callchain) symbol_conf.use_callchain = true;else
tools/perf/util/session.c +38 -0
@@ -12,6 +12,7 @@#include "session.h" #include "sort.h" #include "util.h" +#include "cpumap.h" static int perf_session__open (struct perf_session *self,bool force) {
@@ -1282,3 +1283,40 @@ void perf_session__print_ip(union perf_event *event, } } } + +int perf_session__cpu_bitmap (struct perf_session *session, +const char *cpu_list,unsigned long *cpu_bitmap) +{ +int i; +struct cpu_map *map; + +for (i =0 ; i <PERF_TYPE_MAX ; ++i) { +struct perf_evsel *evsel; + + evsel =perf_session__find_first_evtype (session, i); +if (!evsel) +continue ; + +if (!(evsel->attr.sample_type &PERF_SAMPLE_CPU )) { +pr_err ("File does not contain CPU events. " +"Remove -c option to proceed.\n" ); +return -1 ; + } + } + + map =cpu_map__new (cpu_list); + +for (i =0 ; i < map->nr; i++) { +int cpu = map->map[i]; + +if (cpu >=MAX_NR_CPUS ) { +pr_err ("Requested CPU %d too large. " +"Consider raising MAX_NR_CPUS\n" , cpu); +return -1 ; + } + +set_bit (cpu, cpu_bitmap); + } + +return 0 ; +}
tools/perf/util/session.h +3 -0
@@ -172,4 +172,7 @@ void perf_session__print_ip(union perf_event *event,struct perf_session *session, int print_sym, int print_dso); +int perf_session__cpu_bitmap (struct perf_session *session, +const char *cpu_list,unsigned long *cpu_bitmap); +#endif /* __PERF_SESSION_H */