perf,core: allow invalid context events to be part of sw/hw groups
Subsystems (2)
PATCH 1/5 perf,core: allow invalid context events to be part of sw/hw groups
From: Kan Liang <[email protected]>
The pmu marked as perf_invalid_context don't have any state to switch on context switch. Everything is global. So it is OK to be part of sw/hw groups.
In sched_out/sched_in, del/add must be called, so the perf_invalid_context event can be disabled/enabled accordingly during context switch. The event count only be read when the event is already sched_in.
However group read doesn't work with mix events.
For example,
perf record -e '{cycles,uncore_imc_0/cas_count_read/}:S' -a sleep 1 It always gets EINVAL.
This patch set intends to fix this issue.
perf record -e '{cycles,uncore_imc_0/cas_count_read/}:S' -a sleep 1 [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.202 MB perf.data (12 samples) ]
This patch special case invalid context events and allow them to be part of sw/hw groups.
--- include/linux/perf_event.h | 8 ++++++ kernel/events/core.c | 72 ++++++++++++++++++++++++++++++++++------------ 2 files changed, 61 insertions(+), 19 deletions(-)
include/linux/perf_event.h +8 -0
@@ -711,6 +711,14 @@ static inline bool is_sampling_event(struct perf_event *event)/* * Return 1 for a software event, 0 for a hardware event */ +static inline int is_invalid_context_event (struct perf_event *event) +{ +return event->pmu->task_ctx_nr == perf_invalid_context; +} + +/* +* Return 1 for a software event, 0 for a hardware event +*/ static inline int is_software_event (struct perf_event *event) {return event->pmu->task_ctx_nr == perf_sw_context;
kernel/events/core.c +53 -19
91 lines not on this page · raw message
Signature · 1 line
1.8.3.1
PATCH 2/5 perf,tools: check and re-organize evsel cpu maps
From: Kan Liang <[email protected]>
With the patch 1/5, it's possible to group read events from different pmus. "-C" can be used to set cpu list. The cpu list may be incompatible with pmu's cpumask.
This patch checks the event's cpu maps, and discard the incompatible cpu maps.
event's cpu maps is saved in evsel->cpus during option parse. Then the evlist's cpu maps is created in perf_evlist__create_maps. So the cpu maps can be check and re-organized in perf_evlist__create_maps. Only cpu_list need to check the cpu maps.
--- tools/perf/builtin-top.c | 6 ++-- tools/perf/util/evlist.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 75 insertions(+), 3 deletions(-)
tools/perf/builtin-top.c +3 -3
@@ -1218,9 +1218,6 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)if (target__none (target)) target->system_wide = true; -if (perf_evlist__create_maps (top.evlist, target) <0 ) -usage_with_options (top_usage, options); -if (!top.evlist->nr_entries &&perf_evlist__add_default (top.evlist) <0 ) {ui__error ("Not enough memory for event selector list\n" );
@@ -1229,6 +1226,9 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused) symbol_conf.nr_events = top.evlist->nr_entries; +if (perf_evlist__create_maps (top.evlist, target) <0 ) +usage_with_options (top_usage, options); +if (top.delay_secs <1 ) top.delay_secs =1 ;
tools/perf/util/evlist.c +72 -0
84 lines not on this page · raw message
Signature · 1 line
1.8.3.1
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 2/5] perf,tools: check and re-organize evsel cpu maps
Em Tue, Mar 03, 2015 at 03:54:43AM -0500, [email protected] escreveu:
Quoted 11 lines
From: Kan Liang <[email protected]>
With the patch 1/5, it's possible to group read events from different pmus. "-C" can be used to set cpu list. The cpu list may be incompatible with pmu's cpumask.
This patch checks the event's cpu maps, and discard the incompatible cpu maps.
event's cpu maps is saved inevsel->cpusduring option parse. Then the evlist's cpu maps is created inperf_evlist__create_maps. So the cpu maps can be check and re-organized inperf_evlist__create_maps. Onlycpu_listneed to check the cpu maps.
Humm, I had something done in this area...
Stephane complained about the confusion about which cpumap to use with pmus, so I wrote a patch and sent an RFC, which I think I got no comments, lemme dig it...
- Arnaldo
Quoted 122 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-top.c | 6 ++-- tools/perf/util/evlist.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 75 insertions(+), 3 deletions(-)tools/perf/builtin-top.c +3
-3@@ -1218,9 +1218,6 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)if (target__none (target)) target->system_wide = true; -if (perf_evlist__create_maps (top.evlist, target) <0 ) -usage_with_options (top_usage, options); -if (!top.evlist->nr_entries &&perf_evlist__add_default (top.evlist) <0 ) {ui__error ("Not enough memory for event selector list\n" );@@ -1229,6 +1226,9 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused) symbol_conf.nr_events = top.evlist->nr_entries; +if (perf_evlist__create_maps (top.evlist, target) <0 ) +usage_with_options (top_usage, options); +if (top.delay_secs <1 ) top.delay_secs =1 ;tools/perf/util/evlist.c +72
-0@@ -1026,6 +1026,74 @@ int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,return perf_evlist__mmap_per_cpu (evlist, &mp); } +static int cmp_ids (const void *a,const void *b) +{ +return *(int *)a - *(int *)b; +} + +/* +* Check evsel cpu map according to pmu cpumask and input +* Only available cpu can be stored in evsel->cpus->map. +*/ +static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist) +{ +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +struct perf_evsel *evsel; +int i, j, cpu_nr, tmp; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids); + +evlist__for_each (evlist, evsel) { +if (!evsel->cpus) +continue ; + + cpu_nr =0 ; + j =0 ; +for (i =0 ; i <cpu_map__nr (evsel->cpus);) { + +if (j >= ncpus) { + evsel->cpus->map[i++] =-1 ; +continue ; + } +for (; j < ncpus; j++) { +if (cpus->map[j] < evsel->cpus->map[i]) +continue ; +if (cpus->map[j] == evsel->cpus->map[i]) { + cpu_nr++; + j++; + i++; + }else + evsel->cpus->map[i++] =-1 ; +break ; + } + } + +if (cpu_nr ==cpu_map__nr (evsel->cpus)) +continue ; +if (cpu_nr ==0 ) { +perror ("failed to create CPUs map, please check cpumask" ); +return -1 ; + } + + tmp =0 ; +for (i =0 ; i < cpu_nr; i++) { +if (evsel->cpus->map[i] ==-1 ) { +while (evsel->cpus->map[tmp] ==-1 ) { + tmp++; +BUG_ON (tmp >=cpu_map__nr (evsel->cpus)); + } + evsel->cpus->map[i] = evsel->cpus->map[tmp]; + evsel->cpus->map[tmp] =-1 ; + } + tmp++; + } + evsel->cpus->nr = cpu_nr; + } +return 0 ; +} +int perf_evlist__create_maps (struct perf_evlist *evlist,struct target *target) { evlist->threads = thread_map__new_str(target->pid, target->tid,@@ -1042,6 +1110,10 @@ int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)if (evlist->cpus ==NULL ) goto out_delete_threads; +if (target->cpu_list && + (perf_evlist__check_evsel_cpus (evlist) <0 )) + goto out_delete_threads; +return 0 ; out_delete_threads:Signature · 1 line
1.8.3.1
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 2/5] perf,tools: check and re-organize evsel cpu maps
Em Tue, Mar 03, 2015 at 01:09:29PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 13 lines
Em Tue, Mar 03, 2015 at 03:54:43AM -0500, [email protected] escreveu:
Quoted 11 lines
From: Kan Liang <[email protected]>
With the patch 1/5, it's possible to group read events from different pmus. "-C" can be used to set cpu list. The cpu list may be incompatible with pmu's cpumask.
This patch checks the event's cpu maps, and discard the incompatible cpu maps.
event's cpu maps is saved inevsel->cpusduring option parse. Then the evlist's cpu maps is created inperf_evlist__create_maps. So the cpu maps can be check and re-organized inperf_evlist__create_maps. Onlycpu_listneed to check the cpu maps.
Quoted 5 lines
Humm, I had something done in this area...
Stephane complained about the confusion about which cpumap to use with pmus, so I wrote a patch and sent an RFC, which I think I got no comments, lemme dig it...
Here it is, can you take a look? Stephane?
- Arnaldo
commit 9ecdd9b9bf0a7fd5645957ba4e6a98b6ee526109
Author: Arnaldo Carvalho de Melo <[email protected]> Date: Mon Jan 26 13:43:42 2015 -0300
perf evsel: Set evsel->cpus to the evlist->cpus when not constrained
So that we don't need to know about the evlist all the time and can cope with cases where evsel->cpus were set because it was for an event on a PMU with a cpumask.
echo Link: http://lkml.kernel.org/n/tip-ranpwd -l 24@git.kernel.org
tools/perf/builtin-stat.c +13 -23
@@ -163,16 +163,6 @@ static inline void diff_timespec(struct timespec *r, struct timespec *a, } } -static inline struct cpu_map *perf_evsel__cpus (struct perf_evsel *evsel) -{ -return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus; -} - -static inline int perf_evsel__nr_cpus (struct perf_evsel *evsel) -{ -return perf_evsel__cpus (evsel)->nr; -} -static void perf_evsel__reset_stat_priv (struct perf_evsel *evsel) {int i;
@@ -202,7 +192,7 @@ static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)size_t sz; sz =sizeof (*evsel->counts) + - (perf_evsel__nr_cpus (evsel) *sizeof (struct perf_counts_values )); + (cpu_map__nr (evsel->cpus) *sizeof (struct perf_counts_values )); addr =zalloc (sz);if (!addr)
@@ -235,7 +225,7 @@ static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)evlist__for_each (evlist, evsel) {if (perf_evsel__alloc_stat_priv (evsel) <0 || -perf_evsel__alloc_counts (evsel,perf_evsel__nr_cpus (evsel)) <0 || +perf_evsel__alloc_counts (evsel,cpu_map__nr (evsel->cpus)) <0 || (alloc_raw &&perf_evsel__alloc_prev_raw_counts (evsel) <0 )) goto out_free; }
@@ -269,7 +259,7 @@ static void perf_stat__reset_stats(struct perf_evlist *evlist)evlist__for_each (evlist, evsel) {perf_evsel__reset_stat_priv (evsel); -perf_evsel__reset_counts (evsel,perf_evsel__nr_cpus (evsel)); +perf_evsel__reset_counts (evsel,cpu_map__nr (evsel->cpus)); }memset (runtime_nsecs_stats,0 ,sizeof (runtime_nsecs_stats));
@@ -302,7 +292,7 @@ static int create_perf_stat_counter(struct perf_evsel *evsel) attr->inherit = !no_inherit;if (target__has_cpu (&target)) -return perf_evsel__open_per_cpu (evsel,perf_evsel__cpus (evsel)); +return perf_evsel__open_per_cpu (evsel, evsel->cpus);if (!target__has_task (&target) &&perf_evsel__is_group_leader (evsel)) { attr->disabled =1 ;
@@ -397,7 +387,7 @@ static void zero_per_pkg(struct perf_evsel *counter)static int check_per_pkg (struct perf_evsel *counter,int cpu,bool *skip) {unsigned long *mask = counter->per_pkg_mask; -struct cpu_map *cpus =perf_evsel__cpus (counter); +struct cpu_map *cpus = counter->cpus;int s; *skip = false;
@@ -507,7 +497,7 @@ static int read_counter_aggr(struct perf_evsel *counter)static int read_counter (struct perf_evsel *counter) {int nthreads =thread_map__nr (evsel_list->threads); -int ncpus =perf_evsel__nr_cpus (counter); +int ncpus =cpu_map__nr (counter->cpus);int cpu, thread;if (counter->system_wide)
@@ -727,13 +717,13 @@ static int __run_perf_stat(int argc, const char **argv)if (aggr_mode ==AGGR_GLOBAL ) {evlist__for_each (evsel_list, counter) {read_counter_aggr (counter); -perf_evsel__close_fd (counter,perf_evsel__nr_cpus (counter), +perf_evsel__close_fd (counter,cpu_map__nr (counter->cpus),thread_map__nr (evsel_list->threads)); } }else {evlist__for_each (evsel_list, counter) {read_counter (counter); -perf_evsel__close_fd (counter,perf_evsel__nr_cpus (counter),1 ); +perf_evsel__close_fd (counter,cpu_map__nr (counter->cpus),1 ); } }
@@ -812,7 +802,7 @@ static void aggr_printout(struct perf_evsel *evsel, int id, int nr)case AGGR_NONE :fprintf (output,"CPU%*d%s" , csv_output ?0 :-4 , -perf_evsel__cpus (evsel)->map[id], csv_sep); + evsel->cpus->map[id], csv_sep);break ;case AGGR_GLOBAL :default :
@@ -1216,8 +1206,8 @@ static void print_aggr(char *prefix)evlist__for_each (evsel_list, counter) { val = ena = run =0 ; nr =0 ; -for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { - cpu2 =perf_evsel__cpus (counter)->map[cpu]; +for (cpu =0 ; cpu <cpu_map__nr (counter->cpus); cpu++) { + cpu2 = counter->cpus->map[cpu]; s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
@@ -1339,7 +1329,7 @@ static void print_counter(struct perf_evsel *counter, char *prefix)double uval;int cpu; -for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { +for (cpu =0 ; cpu <cpu_map__nr (counter->cpus); cpu++) { val = counter->counts->cpu[cpu].val; ena = counter->counts->cpu[cpu].ena; run = counter->counts->cpu[cpu].run;
@@ -1350,7 +1340,7 @@ static void print_counter(struct perf_evsel *counter, char *prefix)if (run ==0 || ena ==0 ) {fprintf (output,"CPU%*d%s%*s%s" , csv_output ?0 :-4 , -perf_evsel__cpus (counter)->map[cpu], csv_sep, + counter->cpus->map[cpu], csv_sep, csv_output ?0 :18 , counter->supported ?CNTR_NOT_COUNTED :CNTR_NOT_SUPPORTED , csv_sep);
tools/perf/util/evlist.c +53 -8
117 lines not on this page · raw message
RE: [PATCH 2/5] perf,tools: check and re-organize evsel cpu maps
Quoted 17 lines
Em Tue, Mar 03, 2015 at 01:09:29PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 13 lines
Em Tue, Mar 03, 2015 at 03:54:43AM -0500, [email protected] escreveu:
Quoted 11 lines
From: Kan Liang <[email protected]>
With the patch 1/5, it's possible to group read events from different pmus. "-C" can be used to set cpu list. The cpu list may be incompatible with pmu's cpumask.
This patch checks the event's cpu maps, and discard the incompatible cpu maps.
event's cpu maps is saved inevsel->cpusduring option parse. Then the evlist's cpu maps is created inperf_evlist__create_maps. So the cpu maps can be check and re-organized inperf_evlist__create_maps. Onlycpu_listneed to check the cpu maps.
Quoted 9 lines
Quoted 3 lines
Humm, I had something done in this area...
Stephane complained about the confusion about which cpumap to use
with
Quoted 2 lines
pmus, so I wrote a patch and sent an RFC, which I think I got no comments, lemme dig it...
Here it is, can you take a look? Stephane?
Your patch is more like my 3/5 patch. The difference is your patch force the evsel->cpus = evlist->cpus, if evsel->cpus == NULL.
My patch handle the evsel->cpus == NULL case when using it.
Quoted 11 lines
(quoted diff) +2
-2@@ -1216,8 +1206,8 @@ static void print_aggr(char *prefix) evlist__for_each(evsel_list, counter) { val = ena = run = 0; nr = 0; - for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { - cpu2 = perf_evsel__cpus(counter)- >map[cpu]; + for (cpu = 0; cpu < cpu_map__nr(counter->cpus); cpu++) { + cpu2 = counter->cpus->map[cpu];
tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 1213
Quoted 3 lines
tools/perf/builtin-stat.c +0
-0s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
print_aggr also need to be special handled. In the past, all events use evlist's cpu map,so it uses index to find the real cpu id. Now, event's cpu map are different. The s2 could be wrong. For example, evlist's cpu map is 0,4,5,18. Event's cpu map could be 0,18. When cpu == 1, the return of aggr_get_id must be wrong, since it still use index to find s2.
My 3/5 patch introduce a function perf_evsel__get_cpumap_index to handle it.
Only your patch is not enough, we still need 2/5 and 4/5. 2/5 is used to check if the event's cpu maps are compatible as evlist's cpu map. For example, evlist's cpu map is 1,2,17. Event's cpu map could be 0,18. We can error out earlier.
4/5 is used to special handle the open and mmap. We need to do the same thing as what we did in print_aggr.
Thanks,
Kan
Arnaldo Carvalho de Melo
maintainerRe: [PATCH 2/5] perf,tools: check and re-organize evsel cpu maps
Em Tue, Mar 03, 2015 at 05:09:29PM +0000, Liang, Kan escreveu:
Quoted 24 lines
Quoted 24 lines
Em Tue, Mar 03, 2015 at 01:09:29PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 13 lines
Em Tue, Mar 03, 2015 at 03:54:43AM -0500, [email protected] escreveu:
Quoted 11 lines
From: Kan Liang <[email protected]>
With the patch 1/5, it's possible to group read events from different pmus. "-C" can be used to set cpu list. The cpu list may be incompatible with pmu's cpumask.
This patch checks the event's cpu maps, and discard the incompatible cpu maps.
event's cpu maps is saved inevsel->cpusduring option parse. Then the evlist's cpu maps is created inperf_evlist__create_maps. So the cpu maps can be check and re-organized inperf_evlist__create_maps. Onlycpu_listneed to check the cpu maps.Quoted 3 lines
Humm, I had something done in this area...
Stephane complained about the confusion about which cpumap to use
with
Quoted 2 lines
pmus, so I wrote a patch and sent an RFC, which I think I got no comments, lemme dig it...
Quoted 6 lines
Quoted 2 lines
Here it is, can you take a look? Stephane?
Your patch is more like my 3/5 patch. The difference is your patch force the
evsel->cpus=evlist->cpus, ifevsel->cpus== NULL.
My patch handle theevsel->cpus== NULL case when using it.
Idea is to use evsel->cpus always, not having to special case it and fallback to evlist->cpus, so that we don't have to pass evlist around that often.
Quoted 24 lines
Quoted 11 lines
(quoted diff) +2
-2@@ -1216,8 +1206,8 @@ static void print_aggr(char *prefix) evlist__for_each(evsel_list, counter) { val = ena = run = 0; nr = 0; - for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { - cpu2 = perf_evsel__cpus(counter)- >map[cpu]; + for (cpu = 0; cpu < cpu_map__nr(counter->cpus); cpu++) { + cpu2 = counter->cpus->map[cpu];tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 1213
Quoted 3 lines
tools/perf/builtin-stat.c +0
-0s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
print_aggralso need to be special handled. In the past, all events use evlist's cpu map,so it uses index to find the real cpu id. Now, event's cpu map are different. The s2 could be wrong. For example, evlist's cpu map is 0,4,5,18. Event's cpu map could be 0,18. When cpu == 1, the return ofaggr_get_idmust be wrong, since it still use index to find s2.
My 3/5 patch introduce a functionperf_evsel__get_cpumap_indexto handle it.
Quoted 6 lines
Only your patch is not enough, we still need 2/5 and 4/5. 2/5 is used to check if the event's cpu maps are compatible as evlist's cpu map. For example, evlist's cpu map is 1,2,17. Event's cpu map could be 0,18. We can error out earlier.
4/5 is used to special handle the open and mmap. We need to do the same thing as what we did inprint_aggr.
I'll try to go thru this tomorrow, thanks for checking.
- Arnaldo
RE: [PATCH 2/5] perf,tools: check and re-organize evsel cpu maps
Quoted 65 lines
Em Tue, Mar 03, 2015 at 05:09:29PM +0000, Liang, Kan escreveu:
Quoted 5 lines
Quoted 5 lines
Em Tue, Mar 03, 2015 at 01:09:29PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 1 line
Em Tue, Mar 03, 2015 at 03:54:43AM -0500, [email protected]
escreveu:
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
From: Kan Liang <[email protected]>
With the patch 1/5, it's possible to group read events from different pmus. "-C" can be used to set cpu list. The cpu list may be incompatible with pmu's cpumask.
This patch checks the event's cpu maps, and discard the incompatible cpu maps.
event's cpu maps is saved inevsel->cpusduring option parse. Then the evlist's cpu maps is created inperf_evlist__create_maps. So the cpu maps can be check and re-organized in
perf_evlist__create_maps.Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 1 line
Only
cpu_listneed to check the cpu maps.Humm, I had something done in this area...
Stephane complained about the confusion about which cpumap to
use
Quoted 10 lines
Quoted 6 lines
with
Quoted 2 lines
pmus, so I wrote a patch and sent an RFC, which I think I got no comments, lemme dig it...
Here it is, can you take a look? Stephane?
Your patch is more like my 3/5 patch. The difference is your patch force the
evsel->cpus=evlist->cpus, ifevsel->cpus== NULL. My patch handle theevsel->cpus== NULL case when using it.Idea is to use
evsel->cpusalways, not having to special case it and fallback toevlist->cpus, so that we don't have to pass evlist around that often.Quoted 24 lines
Quoted 11 lines
(quoted diff) +2
-2@@ -1216,8 +1206,8 @@ static void print_aggr(char *prefix) evlist__for_each(evsel_list, counter) { val = ena = run = 0; nr = 0; - for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) { - cpu2 = perf_evsel__cpus(counter)- >map[cpu]; + for (cpu = 0; cpu < cpu_map__nr(counter->cpus); cpu++) { + cpu2 = counter->cpus->map[cpu];tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 1213
Quoted 3 lines
tools/perf/builtin-stat.c +0
-0s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
print_aggralso need to be special handled. In the past, all events use evlist's cpu map,so it uses index to find the real cpu id. Now, event's cpu map are different. The s2 could be wrong. For example, evlist's cpu map is 0,4,5,18. Event's cpu map could be 0,18. When cpu == 1, the return ofaggr_get_idmust be wrong, since it still use index to find s2.
My 3/5 patch introduce a functionperf_evsel__get_cpumap_indexto handle it.
Quoted 8 lines
Quoted 6 lines
Only your patch is not enough, we still need 2/5 and 4/5. 2/5 is used to check if the event's cpu maps are compatible as evlist's cpu map. For example, evlist's cpu map is 1,2,17. Event's cpu map could be 0,18. We can error out earlier.
4/5 is used to special handle the open and mmap. We need to do the same thing as what we did inprint_aggr.I'll try to go thru this tomorrow, thanks for checking.
Hi Arnaldo,
Have you got a chance to check the patch?
Thanks,
Kan
Quoted 2 lines
- Arnaldo
PATCH 3/5 perf,tools: change perf stat to use event's cpu map
In reply to [email protected]From: Kan Liang <[email protected]>
With cpu_list, perf stat should use event's cpu map.
Different events may have different cpu map. So the same index in cpu map does not point to the same CPU id. perf_evsel__get_cpumap_index is introduced to find out the CPU id's index for a given cpu map.
--- tools/perf/builtin-stat.c | 5 +++-- tools/perf/util/evsel.h | 24 ++++++++++++++++++++++++ 2 files changed, 27 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +3 -2
@@ -165,7 +165,7 @@ static inline void diff_timespec(struct timespec *r, struct timespec *a,static inline struct cpu_map *perf_evsel__cpus (struct perf_evsel *evsel) { -return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus; +return evsel->cpus ? evsel->cpus : evsel_list->cpus; }static inline int perf_evsel__nr_cpus (struct perf_evsel *evsel)
@@ -302,7 +302,7 @@ static int create_perf_stat_counter(struct perf_evsel *evsel) attr->inherit = !no_inherit;if (target__has_cpu (&target)) -return perf_evsel__open_per_cpu (evsel,perf_evsel__cpus (evsel)); +return perf_evsel__open_per_cpu (evsel, evsel_list->cpus);if (!target__has_task (&target) &&perf_evsel__is_group_leader (evsel)) { attr->disabled =1 ;
@@ -1218,6 +1218,7 @@ static void print_aggr(char *prefix) nr =0 ;for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { cpu2 =perf_evsel__cpus (counter)->map[cpu]; + cpu2 =perf_evsel__get_cpumap_index (NULL , cpu2, evsel_list->cpus); s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
tools/perf/util/evsel.h +24 -0
@@ -8,6 +8,7 @@#include <linux/types.h> #include "xyarray.h" #include "symbol.h" +#include "cpumap.h" struct perf_counts_values { union {
@@ -359,4 +360,27 @@ static inline bool has_branch_callstack(struct perf_evsel *evsel) {return evsel->attr.branch_sample_type &PERF_SAMPLE_BRANCH_CALL_STACK ; } + +static inline int perf_evsel__get_cpumap_index (struct perf_evsel *evsel, +int cpu, +struct cpu_map *evlist_cpus) +{ +struct cpu_map *cpus; +int i; + +if (evlist_cpus ==NULL ) +return -1 ; + +if (evsel && evsel->cpus) + cpus = evsel->cpus; +else + cpus = evlist_cpus; + +for (i =0 ; i < cpus->nr; i++) { +if (cpu == cpus->map[i]) +return i; + } +return -1 ; +} +#endif /* __PERF_EVSEL_H */
Signature · 1 line
1.8.3.1
PATCH 4/5 perf,tools: open/mmap event according to event's cpu map not evlist's
In reply to [email protected]From: Kan Liang <[email protected]>
perf tool can open/mmap events per cpu, but the cpu list is from evlist's cpu map. That means all events share the same cpu map. However, some events like uncore events have cpu mask. So the global cpu map doesn't work well with mixed pmu events.
In this patch, the event's cpu map will be used to provide the available cpu list for perf open. If it's unavailable, evlist's map will be used. Since different event's cpu list could vary, we cannot rely on the index to get group leader's fd. In get_group_fd, the cpu id is used to get correct fd.perf_evlist__mmap also need to be changed, since Leader's fd has to be mmaped before member's fd. So evlist__for_each must be the outermost of the loop.
--- tools/perf/util/evlist.c | 110 ++++++++++++++++++++++++++++------------------- tools/perf/util/evsel.c | 30 +++++++++---- 2 files changed, 87 insertions(+), 53 deletions(-)
tools/perf/util/evlist.c +65 -45
@@ -792,51 +792,48 @@ static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,return 0 ; } -static int perf_evlist__mmap_per_evsel (struct perf_evlist *evlist,int idx, -struct mmap_params *mp,int cpu, -int thread,int *output) +static int perf_evlist__mmap_per_evsel (struct perf_evlist *evlist, +struct perf_evsel *evsel, +int idx,struct mmap_params *mp, +int cpu,int thread,int *output) { -struct perf_evsel *evsel; +int fd; -evlist__for_each (evlist, evsel) { -int fd; +if (evsel->system_wide && thread) +return 0 ; -if (evsel->system_wide && thread) -continue ; + fd =FD (evsel, cpu, thread); - fd =FD (evsel, cpu, thread); +if (*output ==-1 ) { + *output = fd; +if (__perf_evlist__mmap (evlist, idx, mp, *output) <0 ) +return -1 ; + }else { +if (ioctl (fd,PERF_EVENT_IOC_SET_OUTPUT , *output) !=0 ) +return -1 ; -if (*output ==-1 ) { - *output = fd; -if (__perf_evlist__mmap (evlist, idx, mp, *output) <0 ) -return -1 ; - }else { -if (ioctl (fd,PERF_EVENT_IOC_SET_OUTPUT , *output) !=0 ) -return -1 ; +perf_evlist__mmap_get (evlist, idx); + } -perf_evlist__mmap_get (evlist, idx); - } +/* +* The system_wide flag causes a selected event to be opened +* always without a pid. Consequently it will never get a +* POLLHUP, but it is used for tracking in combination with +* other events, so it should not need to be polled anyway. +* Therefore don't add it for polling. +*/ +if (!evsel->system_wide && +__perf_evlist__add_pollfd (evlist, fd, idx) <0 ) { +perf_evlist__mmap_put (evlist, idx); +return -1 ; + } -/* -* The system_wide flag causes a selected event to be opened -* always without a pid. Consequently it will never get a -* POLLHUP, but it is used for tracking in combination with -* other events, so it should not need to be polled anyway. -* Therefore don't add it for polling. -*/ -if (!evsel->system_wide && -__perf_evlist__add_pollfd (evlist, fd, idx) <0 ) { -perf_evlist__mmap_put (evlist, idx); +if (evsel->attr.read_format &PERF_FORMAT_ID ) { +if (perf_evlist__id_add_fd (evlist, evsel, cpu, thread, + fd) <0 )return -1 ; - } - -if (evsel->attr.read_format &PERF_FORMAT_ID ) { -if (perf_evlist__id_add_fd (evlist, evsel, cpu, thread, - fd) <0 ) -return -1 ; -perf_evlist__set_sid_idx (evlist, evsel, idx, cpu, - thread); - } +perf_evlist__set_sid_idx (evlist, evsel, idx, cpu, + thread); }return 0 ;
@@ -848,23 +845,43 @@ static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,int cpu, thread;int nr_cpus =cpu_map__nr (evlist->cpus);int nr_threads =thread_map__nr (evlist->threads); +int *output =malloc (nr_cpus *sizeof (int )); +struct cpu_map *map; +int evlist_cpu; +struct perf_evsel *evsel;pr_debug2 ("perf event ring buffer mmapped per cpu\n" ); -for (cpu =0 ; cpu < nr_cpus; cpu++) { -int output =-1 ; -for (thread =0 ; thread < nr_threads; thread++) { -if (perf_evlist__mmap_per_evsel (evlist, cpu, mp, cpu, - thread, &output)) +for (cpu =0 ; cpu < nr_cpus; cpu++) + output[cpu] =-1 ; + +evlist__for_each (evlist, evsel) { +if (evsel->cpus) + map = evsel->cpus; +else + map = evlist->cpus; + +for (cpu =0 ; cpu < map->nr; cpu++) { + evlist_cpu =perf_evsel__get_cpumap_index (NULL , map->map[cpu], evlist->cpus); +if (evlist_cpu <0 ) goto out_unmap; + +for (thread =0 ; thread < nr_threads; thread++) { +if (perf_evlist__mmap_per_evsel (evlist, evsel, evlist_cpu, + mp, cpu, thread, + &output[evlist_cpu])) + goto out_unmap; + } } } +free (output);return 0 ; out_unmap:for (cpu =0 ; cpu < nr_cpus; cpu++)__perf_evlist__munmap (evlist, cpu); +free (output);return -1 ; }
@@ -873,14 +890,17 @@ static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, {int thread;int nr_threads =thread_map__nr (evlist->threads); +struct perf_evsel *evsel;pr_debug2 ("perf event ring buffer mmapped per thread\n" );for (thread =0 ; thread < nr_threads; thread++) {int output =-1 ; -if (perf_evlist__mmap_per_evsel (evlist, thread, mp,0 , thread, - &output)) - goto out_unmap; +evlist__for_each (evlist, evsel) { +if (perf_evlist__mmap_per_evsel (evlist, evsel, thread, + mp,0 , thread, &output)) + goto out_unmap; + } }return 0 ;
tools/perf/util/evsel.c +22 -8
@@ -981,10 +981,11 @@ int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,return 0 ; } -static int get_group_fd (struct perf_evsel *evsel,int cpu,int thread) +static int get_group_fd (struct perf_evsel *evsel,int cpu, +int thread,struct cpu_map *cpus) {struct perf_evsel *leader = evsel->leader; -int fd; +int fd, leader_cpu;if (perf_evsel__is_group_leader (evsel))return -1 ;
@@ -995,9 +996,16 @@ static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)*/ BUG_ON (!leader->fd); - fd =FD (leader, cpu, thread); +if (cpu <0 ) + fd =FD (leader,0 , thread); +else { + leader_cpu =perf_evsel__get_cpumap_index (leader, cpu, cpus); +if (leader_cpu >=0 ) + fd =FD (leader, leader_cpu, thread); +else +return -1 ; + }BUG_ON (fd ==-1 ); -return fd; }
@@ -1068,6 +1076,7 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,int cpu, thread, nthreads;unsigned long flags =PERF_FLAG_FD_CLOEXEC ;int pid =-1 , err; +struct cpu_map *cpumap;enum {NO_CHANGE ,SET_TO_MAX ,INCREASED_MAX } set_rlimit =NO_CHANGE ;if (evsel->system_wide)
@@ -1084,6 +1093,11 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus, pid = evsel->cgrp->fd; } +if (evsel->cpus) + cpumap = evsel->cpus; +else + cpumap = cpus; + fallback_missing_features:if (perf_missing_features.cloexec) flags &= ~(unsigned long )PERF_FLAG_FD_CLOEXEC ;
@@ -1098,7 +1112,7 @@ retry_sample_id:if (verbose >=2 )perf_event_attr__fprintf (&evsel->attr, stderr); -for (cpu =0 ; cpu < cpus->nr; cpu++) { +for (cpu =0 ; cpu < cpumap->nr; cpu++) {for (thread =0 ; thread < nthreads; thread++) {int group_fd;
@@ -1106,14 +1120,14 @@ retry_sample_id:if (!evsel->cgrp && !evsel->system_wide) pid = threads->map[thread]; - group_fd =get_group_fd (evsel, cpu, thread); + group_fd =get_group_fd (evsel, cpumap->map[cpu], thread, cpus); retry_open:pr_debug2 ("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n" , - pid, cpus->map[cpu], group_fd, flags); + pid, cpumap->map[cpu], group_fd, flags);FD (evsel, cpu, thread) =sys_perf_event_open (&evsel->attr, pid, - cpus->map[cpu], + cpumap->map[cpu], group_fd, flags);if (FD (evsel, cpu, thread) <0 ) { err = -errno;
Signature · 1 line
1.8.3.1
PATCH 5/5 perf/x86/intel/uncore: do not implicitly set uncore event cpu
In reply to [email protected]From: Kan Liang <[email protected]>
There is cpumask exposed to the uncore pmu sysfs directory. User should set the cpu according to the cpumask. Kernel should not implicitly change the event->cpu.
--- arch/x86/kernel/cpu/perf_event_intel_uncore.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
arch/x86/kernel/cpu/perf_event_intel_uncore.c +1 -2
@@ -621,9 +621,8 @@ static int uncore_pmu_event_init(struct perf_event *event)if (event->cpu <0 )return -EINVAL ; box =uncore_pmu_to_box (pmu, event->cpu); -if (!box || box->cpu <0 ) +if (!box || box->cpu <0 || (box->cpu != event->cpu))return -EINVAL ; - event->cpu = box->cpu; event->hw.idx =-1 ; event->hw.last_tag = ~0ULL ;
Signature · 1 line
1.8.3.1