perf,tools: open event on evsel cpus and threads
PATCH 1/1 perf,tools: open event on evsel cpus and threads
From: Kan Liang <[email protected]>
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)
tools/perf/builtin-record.c +1 -1
@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);
tools/perf/util/evlist.c +4 -0
@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;
Signature · 1 line
1.8.3.1
Quoted 4 lines · On Fri, Aug 21, 2015 at 02:23:14AM -0400, [email protected] wrote:
From: Kan Liang <[email protected]>
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
thanks,
jirka
Quoted 38 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)tools/perf/builtin-record.c +1
-1@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +4
-0@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;Signature · 1 line
1.8.3.1
Arnaldo Carvalho de Melo
maintainerEm Fri, Aug 21, 2015 at 03:54:36PM +0200, Jiri Olsa escreveu:
Quoted 7 lines · On Fri, Aug 21, 2015 at 02:23:14AM -0400, [email protected] wrote:
Quoted 4 lines
From: Kan Liang <[email protected]>
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
Acked-by Jiri Olsa <[email protected]>
Applying, I wonder if this isn't affecting other tools as well... And also perf_evlist__open(), that has to be fixed as well, right?
- Arnaldo
Quoted 41 lines
thanks,
jirkaQuoted 38 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)tools/perf/builtin-record.c +1
-1@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +4
-0@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;Signature · 1 line
1.8.3.1
Quoted 4 lines
Em Fri, Aug 21, 2015 at 03:54:36PM +0200, Jiri Olsa escreveu:
Quoted 1 line · On Fri, Aug 21, 2015 at 02:23:14AM -0400, [email protected] wrote:
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
From: Kan Liang <[email protected]>
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
Acked-by Jiri Olsa <[email protected]>Applying, I wonder if this isn't affecting other tools as well... And also
No, it doesn't affect other tools.
Quoted 2 lines
perf_evlist__open(), that has to be fixed as well, right?
Umm... Right.
We should fix it in perf_evlist__open as well.
IIRC, Jirka once planned to send out a cleanup patch separately for this kind of issue (perf_evsel__enable also need to be fixed) on other tools.
Thanks,
Kan
Quoted 48 lines
- Arnaldo
Quoted 16 lines
thanks,
jirkaQuoted 13 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)(quoted diff) +0
-0
b/tools/perf/builtin-record.cindex 25cf6b4..a0178bf 100644tools/perf/builtin-record.c +0
-0@@ -279,7 +279,7 @@ static int record__open(struct record *rec)Quoted 3 lines
tools/perf/builtin-record.c static int record__open(struct record *rec)
Quoted 3 lines
tools/perf/builtin-record.c +0
-1evlist__for_each (evlist, pos) { try_again: - if (perf_evsel__open(pos, evlist->cpus, evlist->threads) <0) {
Quoted 2 lines
tools/perf/builtin-record.c static int record__open(struct record *rec) · line 283
Quoted 2 lines
tools/perf/builtin-record.c +1
-0+if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) { if (perf_evsel__fallback(pos, errno, msg,sizeof(msg))) {
Quoted 7 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct
Quoted 7 lines
tools/perf/builtin-record.c +0
-0if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +0
-0@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct
perf_evlist*evlist, structperf_evsel**eQuoted 5 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 5 lines
tools/perf/util/evlist.c +2
-0if (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit.
Quoted 3 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 3 lines
tools/perf/util/evlist.c +2
-0+* So evlist and evsel should always be same. +*/ err = perf_evsel__apply_filter(evsel, ncpus, nthreads,
evsel->filter);Quoted 5 lines
Quoted 5 lines
tools/perf/util/evlist.c +0
-1if (err) { *err_evsel = evsel; --1.8.3.1
Arnaldo Carvalho de Melo
maintainerEm Mon, Aug 31, 2015 at 09:06:29PM +0000, Liang, Kan escreveu:
Quoted 15 lines
Quoted 4 lines
Em Fri, Aug 21, 2015 at 03:54:36PM +0200, Jiri Olsa escreveu:
Quoted 1 line · On Fri, Aug 21, 2015 at 02:23:14AM -0400, [email protected] wrote:
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
From: Kan Liang <[email protected]>
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
Acked-by Jiri Olsa <[email protected]>Applying, I wonder if this isn't affecting other tools as well... And also
No, it doesn't affect other tools.
Quoted 8 lines
Quoted 2 lines
perf_evlist__open(), that has to be fixed as well, right?Umm... Right.
We should fix it inperf_evlist__openas well.
IIRC, Jirka once planned to send out a cleanup patch separately for this kind of issue (perf_evsel__enablealso need to be fixed) on other tools.
So now that we have a backpointer, perhaps we should have a:
struct thread_map *perf_evsel__threads(struct perf_evsel *evsel)
{
return evsel->threads ?: evsel->evlist ? evsel->evlist->threads : NULL;
}Ditto for cpus and then change all occurrences of evsel->threads with perf_evsel__threads(), right?
Quoted 51 lines
Thanks,
KanQuoted 48 lines
- Arnaldo
Quoted 16 lines
thanks,
jirkaQuoted 13 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)(quoted diff) +0
-0
b/tools/perf/builtin-record.cindex 25cf6b4..a0178bf 100644tools/perf/builtin-record.c +0
-0@@ -279,7 +279,7 @@ static int record__open(struct record *rec)Quoted 3 lines
tools/perf/builtin-record.c static int record__open(struct record *rec)
Quoted 3 lines
tools/perf/builtin-record.c +0
-1evlist__for_each (evlist, pos) { try_again: - if (perf_evsel__open(pos, evlist->cpus, evlist->threads) <0) {
Quoted 2 lines
tools/perf/builtin-record.c static int record__open(struct record *rec) · line 283
Quoted 2 lines
tools/perf/builtin-record.c +1
-0+if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) { if (perf_evsel__fallback(pos, errno, msg,sizeof(msg))) {
Quoted 7 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct
Quoted 7 lines
tools/perf/builtin-record.c +0
-0if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +0
-0@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct
perf_evlist*evlist, structperf_evsel**eQuoted 5 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 5 lines
tools/perf/util/evlist.c +2
-0if (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit.
Quoted 3 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 3 lines
tools/perf/util/evlist.c +2
-0+* So evlist and evsel should always be same. +*/ err = perf_evsel__apply_filter(evsel, ncpus, nthreads,
evsel->filter);Quoted 5 lines
Quoted 5 lines
tools/perf/util/evlist.c +0
-1if (err) { *err_evsel = evsel; --1.8.3.1
Quoted 30 lines
Em Mon, Aug 31, 2015 at 09:06:29PM +0000, Liang, Kan escreveu:
Quoted 16 lines
Quoted 6 lines
Em Fri, Aug 21, 2015 at 03:54:36PM +0200, Jiri Olsa escreveu:
Quoted 3 lines · On Fri, Aug 21, 2015 at 02:23:14AM -0400, [email protected] wrote:
Quoted 2 lines
From: Kan Liang <[email protected]>
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
Acked-by Jiri Olsa <[email protected]>Applying, I wonder if this isn't affecting other tools as well... And also
No, it doesn't affect other tools.
Quoted 8 lines
Quoted 2 lines
perf_evlist__open(), that has to be fixed as well, right?Umm... Right.
We should fix it inperf_evlist__openas well.
IIRC, Jirka once planned to send out a cleanup patch separately for this kind of issue (perf_evsel__enablealso need to be fixed) on other tools.
Quoted 8 lines
So now that we have a backpointer, perhaps we should have a:
struct thread_map *perf_evsel__threads(struct perf_evsel *evsel) { return evsel->threads ?: evsel->evlist ? evsel->evlist->threads :NULL; }
Ditto for cpus and then change all occurrences of
evsel->threadswithperf_evsel__threads(), right?
No. We need to change all evlist->threads with evsel->threads. Ditto for cpus.
The event should use its own thread/cpu list.
Quoted 56 lines
Quoted 39 lines
Thanks,
KanQuoted 18 lines
- Arnaldo
Quoted 16 lines
thanks,
jirkaQuoted 13 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)(quoted diff) +0
-0
b/tools/perf/builtin-record.cindex 25cf6b4..a0178bf 100644tools/perf/builtin-record.c +0
-0@@ -279,7 +279,7 @@ static int record__open(struct record *rec)Quoted 3 lines
Quoted 3 lines
tools/perf/builtin-record.c static int record__open(struct record *rec)
Quoted 3 lines
tools/perf/builtin-record.c +0
-1evlist__for_each (evlist, pos) { try_again: - if (perf_evsel__open(pos, evlist->cpus, evlist-threads) <
Quoted 2 lines
0) {
Quoted 1 line
tools/perf/builtin-record.c static int record__open(struct record *rec) · line 282
Quoted 1 line
tools/perf/builtin-record.c +1
-0+ if (perf_evsel__open(pos, pos->cpus, pos-threads) < 0) {
Quoted 10 lines
Quoted 1 line
tools/perf/builtin-record.c static int record__open(struct record *rec) · line 283
Quoted 1 line
tools/perf/builtin-record.c +0
-0if (perf_evsel__fallback(pos, errno, msg,sizeof(msg))) {
Quoted 7 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct
Quoted 7 lines
tools/perf/builtin-record.c +0
-0if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +0
-0@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct
perf_evlist*evlist, structperf_evsel**eQuoted 5 lines
Quoted 5 lines
Quoted 5 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 5 lines
tools/perf/util/evlist.c +2
-0if (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't
Quoted 4 lines
Quoted 4 lines
have cpu limit.
Quoted 3 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 3 lines
tools/perf/util/evlist.c +2
-0+* So evlist and evsel should always be same. +*/ err = perf_evsel__apply_filter(evsel, ncpus,nthreads,
Quoted 6 lines
Quoted 6 lines
evsel->filter);Quoted 5 lines
Quoted 5 lines
tools/perf/util/evlist.c +0
-1if (err) { *err_evsel = evsel; --1.8.3.1
Arnaldo Carvalho de Melo
maintainerEm Mon, Aug 31, 2015 at 09:28:43PM +0000, Liang, Kan escreveu:
Quoted 36 lines
Quoted 36 lines
Em Mon, Aug 31, 2015 at 09:06:29PM +0000, Liang, Kan escreveu:
Quoted 16 lines
Quoted 6 lines
Em Fri, Aug 21, 2015 at 03:54:36PM +0200, Jiri Olsa escreveu:
Quoted 3 lines · On Fri, Aug 21, 2015 at 02:23:14AM -0400, [email protected] wrote:
Quoted 2 lines
From: Kan Liang <[email protected]>
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
evsel may have different cpus and threads as evlist's. Use it's own cpus and threads, when open evsel in perf record.
Acked-by Jiri Olsa <[email protected]>Applying, I wonder if this isn't affecting other tools as well... And also
No, it doesn't affect other tools.
Quoted 8 lines
Quoted 2 lines
perf_evlist__open(), that has to be fixed as well, right?Umm... Right.
We should fix it inperf_evlist__openas well.
IIRC, Jirka once planned to send out a cleanup patch separately for this kind of issue (perf_evsel__enablealso need to be fixed) on other tools.So now that we have a backpointer, perhaps we should have a:
struct thread_map *perf_evsel__threads(struct perf_evsel *evsel) { return evsel->threads ?: evsel->evlist ? evsel->evlist->threads :NULL; }
Quoted 6 lines
Quoted 2 lines
Ditto for cpus and then change all occurrences of
evsel->threadswithperf_evsel__threads(), right?No. We need to change all
evlist->threadswithevsel->threads. Ditto for cpus.
The event should use its own thread/cpu list.
Which is equivalent, but IIRC doing it the way you do is better because we do that forwarding already, right?
Lemme check... Yeah, it is done in perf_evlist__propagate_maps(), so, as long as that function was called, we can access evsel->{threads,cpus} and it will use evlist->{threads,cpus} if those entries were not set before perf_evlist__propagate_maps() was called.
- Arnaldo
Quoted 55 lines
Quoted 55 lines
Quoted 38 lines
Thanks,
KanQuoted 18 lines
- Arnaldo
Quoted 16 lines
thanks,
jirkaQuoted 13 lines
Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)(quoted diff) +0
-0
b/tools/perf/builtin-record.cindex 25cf6b4..a0178bf 100644tools/perf/builtin-record.c +0
-0@@ -279,7 +279,7 @@ static int record__open(struct record *rec)Quoted 3 lines
Quoted 3 lines
tools/perf/builtin-record.c static int record__open(struct record *rec)
Quoted 3 lines
tools/perf/builtin-record.c +0
-1evlist__for_each (evlist, pos) { try_again: - if (perf_evsel__open(pos, evlist->cpus, evlist-threads) <
Quoted 2 lines
0) {
Quoted 1 line
tools/perf/builtin-record.c static int record__open(struct record *rec) · line 282
Quoted 1 line
tools/perf/builtin-record.c +1
-0+ if (perf_evsel__open(pos, pos->cpus, pos-threads) < 0) {
Quoted 10 lines
Quoted 1 line
tools/perf/builtin-record.c static int record__open(struct record *rec) · line 283
Quoted 1 line
tools/perf/builtin-record.c +0
-0if (perf_evsel__fallback(pos, errno, msg,sizeof(msg))) {
Quoted 7 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct
Quoted 7 lines
tools/perf/builtin-record.c +0
-0if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +0
-0@@ -1179,6 +1179,10 @@ int perf_evlist__apply_filters(struct
perf_evlist*evlist, structperf_evsel**eQuoted 5 lines
Quoted 5 lines
Quoted 5 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 5 lines
tools/perf/util/evlist.c +2
-0if (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't
Quoted 4 lines
Quoted 4 lines
have cpu limit.
Quoted 3 lines
tools/perf/util/evlist.c int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **e
Quoted 3 lines
tools/perf/util/evlist.c +2
-0+* So evlist and evsel should always be same. +*/ err = perf_evsel__apply_filter(evsel, ncpus,nthreads,
Quoted 6 lines
Quoted 6 lines
evsel->filter);Quoted 5 lines
Quoted 5 lines
tools/perf/util/evlist.c +0
-1if (err) { *err_evsel = evsel; --1.8.3.1
tip-bot for Kan Liang
merge bot[tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
In reply to [email protected]perf evlist: Open event on evsel cpus and threads
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300
perf evlist: Open event on evsel cpus and threads
An evsel may have different cpus and threads than the evlist it is in.
Use it's own cpus and threads, when opening the evsel in 'perf record'.
--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)
tools/perf/builtin-record.c +1 -1
@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);
tools/perf/util/evlist.c +4 -0
@@ -1181,6 +1181,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;
Adrian Hunter
reviewerRe: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Quoted 9 lines · tip-bot for Kan Liang
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300perf evlist: Open event on evsel cpus and threads
Quoted 8 lines
An evsel may have different cpus and threads than the evlist it is in.
Use it's own cpus and threads, when opening the evsel in 'perf record'.
Signed-off-by Kan Liang <[email protected]>Cc Jiri Olsa <[email protected]>Signed-off-by Arnaldo Carvalho de Melo <[email protected]>
Just noticed this breaks Intel PT. Will have to investigate further.
Quoted 34 lines
--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)tools/perf/builtin-record.c +1
-1@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +4
-0@@ -1181,6 +1181,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;
Re: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 12 lines · tip-bot for Kan Liang
Quoted 11 lines
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300perf evlist: Open event on evsel cpus and threads
An evsel may have different cpus and threads than the evlist it is in.
Quoted 8 lines
Quoted 6 lines
Use it's own cpus and threads, when opening the evsel in 'perf record'.
Signed-off-by Kan Liang <[email protected]>Cc Jiri Olsa <[email protected]>Signed-off-by Arnaldo Carvalho de Melo <[email protected]>Just noticed this breaks Intel PT. Will have to investigate further.
What kind of breakage?
It all should be equivalent to before, its just that it uses evsel->{threads,cpus} while before it was using evlist->{threads,cpus}, but that should point to the same thing if that perf_evlist__propagate_maps() method was called, so I assume this is some segfault?
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
- Arnaldo
Quoted 34 lines
Quoted 34 lines
--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)tools/perf/builtin-record.c +1
-1@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +4
-0@@ -1181,6 +1181,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;
Adrian Hunter
reviewerRe: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Quoted 15 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 14 lines · tip-bot for Kan Liang
Quoted 13 lines
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300perf evlist: Open event on evsel cpus and threads
An evsel may have different cpus and threads than the evlist it is in.
Use it's own cpus and threads, when opening the evsel in 'perf record'.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
Signed-off-by Kan Liang <[email protected]>Cc Jiri Olsa <[email protected]>Signed-off-by Arnaldo Carvalho de Melo <[email protected]>Just noticed this breaks Intel PT. Will have to investigate further.
What kind of breakage?
Can't open the sched_switch event
Quoted 6 lines
It all should be equivalent to before, its just that it uses evsel->{threads,cpus} while before it was using evlist->{threads,cpus}, but that should point to the same thing if that
perf_evlist__propagate_maps()method was called, so I assume this is some segfault?
I think maybe it doesn't consider evsels added later
Quoted 4 lines
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
There is "Test tracking with sched_switch" but you need to expose it to the same issue i.e.
tools/perf/util/evlist.c +1 -1
@@ -1433,7 +1433,7 @@ int perf_evlist__open(struct perf_evlist *evlist)perf_evlist__update_id_pos (evlist);evlist__for_each (evlist, evsel) { - err =perf_evsel__open (evsel, evlist->cpus, evlist->threads); + err =perf_evsel__open (evsel, evsel->cpus, evsel->threads);if (err <0 ) goto out_err; }
Quoted 37 lines
- Arnaldo
Quoted 34 lines
Quoted 34 lines
--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)tools/perf/builtin-record.c +1
-1@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +4
-0@@ -1181,6 +1181,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;
Re: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Em Thu, Sep 03, 2015 at 07:23:43PM +0300, Adrian Hunter escreveu:
Quoted 28 lines · Arnaldo Carvalho de Melo
Quoted 15 lines
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 14 lines · tip-bot for Kan Liang
Quoted 13 lines
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300perf evlist: Open event on evsel cpus and threads
An evsel may have different cpus and threads than the evlist it is in.
Use it's own cpus and threads, when opening the evsel in 'perf record'.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
Signed-off-by Kan Liang <[email protected]>Cc Jiri Olsa <[email protected]>Signed-off-by Arnaldo Carvalho de Melo <[email protected]>Just noticed this breaks Intel PT. Will have to investigate further.
What kind of breakage?
Can't open the
sched_switcheventQuoted 1 line
Quoted 7 lines
Quoted 5 lines
It all should be equivalent to before, its just that it uses evsel->{threads,cpus} while before it was using evlist->{threads,cpus}, but that should point to the same thing if that
perf_evlist__propagate_maps()method was called, so I assume this is some segfault?I think maybe it doesn't consider evsels added later
A-ha, so when using perf_evlist__add() we need to do this, i.e. if evsel->{threads||cpus} is not set, fill it in with the evlist-> member?
Quoted 7 lines
Quoted 4 lines
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
There is "Test tracking with
sched_switch" but you need to expose it to the same issue i.e.
Sure, Kan and Jiri were talking about the need to go doing these changes, Jiri? Kan?
- Arnaldo
Quoted 53 lines
tools/perf/util/evlist.c +1
-1@@ -1433,7 +1433,7 @@ int perf_evlist__open(struct perf_evlist *evlist)perf_evlist__update_id_pos (evlist);evlist__for_each (evlist, evsel) { - err =perf_evsel__open (evsel, evlist->cpus, evlist->threads); + err =perf_evsel__open (evsel, evsel->cpus, evsel->threads);if (err <0 ) goto out_err; }Quoted 40 lines
- Arnaldo
Quoted 34 lines
Quoted 34 lines
--- tools/perf/builtin-record.c | 2 +- tools/perf/util/evlist.c | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)tools/perf/builtin-record.c +1
-1@@ -279,7 +279,7 @@ static int record__open(struct record *rec)evlist__for_each (evlist, pos) { try_again: -if (perf_evsel__open (pos, evlist->cpus, evlist->threads) <0 ) { +if (perf_evsel__open (pos, pos->cpus, pos->threads) <0 ) {if (perf_evsel__fallback (pos, errno, msg,sizeof (msg))) {if (verbose)ui__warning ("%s\n" , msg);tools/perf/util/evlist.c +4
-0@@ -1181,6 +1181,10 @@ int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **eif (evsel->filter ==NULL )continue ; +/* +* filters only work for tracepoint event, which doesn't have cpu limit. +* So evlist and evsel should always be same. +*/ err =perf_evsel__apply_filter (evsel, ncpus, nthreads, evsel->filter);if (err) { *err_evsel = evsel;
Re: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Quoted 45 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 07:23:43PM +0300, Adrian Hunter escreveu:
Quoted 28 lines · Arnaldo Carvalho de Melo
Quoted 15 lines
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 14 lines · tip-bot for Kan Liang
Quoted 13 lines
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300perf evlist: Open event on evsel cpus and threads
An evsel may have different cpus and threads than the evlist it is in.
Use it's own cpus and threads, when opening the evsel in 'perf record'.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
Signed-off-by Kan Liang <[email protected]>Cc Jiri Olsa <[email protected]>Signed-off-by Arnaldo Carvalho de Melo <[email protected]>Just noticed this breaks Intel PT. Will have to investigate further.
What kind of breakage?
Can't open the
sched_switcheventQuoted 1 line
Quoted 7 lines
Quoted 5 lines
It all should be equivalent to before, its just that it uses evsel->{threads,cpus} while before it was using evlist->{threads,cpus}, but that should point to the same thing if that
perf_evlist__propagate_maps()method was called, so I assume this is some segfault?I think maybe it doesn't consider evsels added later
A-ha, so when using
perf_evlist__add()we need to do this, i.e. if evsel->{threads||cpus} is not set, fill it in with the evlist-> member?Quoted 5 lines
Quoted 4 lines
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
Quoted 6 lines
Quoted 2 lines
There is "Test tracking with
sched_switch" but you need to expose it to the same issue i.e.Sure, Kan and Jiri were talking about the need to go doing these changes, Jiri? Kan?
perf_evlist__propagate_maps is called from perf_evlist__create_maps, so if evsel is added later it will not be affected, perhaps we need something like below
jirka
---
tools/perf/util/evlist.c +3 -0
@@ -133,6 +133,9 @@ void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)if (!evlist->nr_entries++)perf_evlist__set_id_pos (evlist); + + entry->cpus =cpu_map__get (evlist->cpus); + entry->threads =thread_map__get (evlist->threads); } void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
Adrian Hunter
reviewerRe: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Quoted 49 lines · Jiri Olsa
Quoted 48 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 07:23:43PM +0300, Adrian Hunter escreveu:
Quoted 28 lines · Arnaldo Carvalho de Melo
Quoted 15 lines
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 14 lines · tip-bot for Kan Liang
Quoted 13 lines
Commit-ID: d988d5ee647861706bc7a391ddbc29429b50f00e Gitweb: http://git.kernel.org/tip/d988d5ee647861706bc7a391ddbc29429b50f00e Author: Kan Liang <[email protected]> AuthorDate: Fri, 21 Aug 2015 02:23:14 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 31 Aug 2015 17:28:01 -0300perf evlist: Open event on evsel cpus and threads
An evsel may have different cpus and threads than the evlist it is in.
Use it's own cpus and threads, when opening the evsel in 'perf record'.
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
Signed-off-by Kan Liang <[email protected]>Cc Jiri Olsa <[email protected]>Signed-off-by Arnaldo Carvalho de Melo <[email protected]>Just noticed this breaks Intel PT. Will have to investigate further.
What kind of breakage?
Can't open the
sched_switcheventQuoted 1 line
Quoted 7 lines
Quoted 5 lines
It all should be equivalent to before, its just that it uses evsel->{threads,cpus} while before it was using evlist->{threads,cpus}, but that should point to the same thing if that
perf_evlist__propagate_maps()method was called, so I assume this is some segfault?I think maybe it doesn't consider evsels added later
A-ha, so when using
perf_evlist__add()we need to do this, i.e. if evsel->{threads||cpus} is not set, fill it in with the evlist-> member?Quoted 7 lines
Quoted 4 lines
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
There is "Test tracking with
sched_switch" but you need to expose it to the same issue i.e.
Quoted 7 lines
Quoted 3 lines
Sure, Kan and Jiri were talking about the need to go doing these changes, Jiri? Kan?
perf_evlist__propagate_mapsis called fromperf_evlist__create_maps, so if evsel is added later it will not be affected, perhaps we need something like below
Yes but it would be nice to have a single function to do it that knows the rules for when the evsel->cpus should be retained. Say:
int __perf_evlist__propagate_maps(struct perf_evlist *evlist, struct perf_evsel *evsel) then call it from perf_evlist__propagate_maps(), perf_evlist__add() and perf_evlist__splice_list_tail(). Probably need to add a member to perf_evlist to know when the evlist->cpus should replace the evsel->cpus
Quoted 20 lines
jirka
---
tools/perf/util/evlist.c +3
-0@@ -133,6 +133,9 @@ void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)if (!evlist->nr_entries++)perf_evlist__set_id_pos (evlist); + + entry->cpus =cpu_map__get (evlist->cpus); + entry->threads =thread_map__get (evlist->threads); } void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
Re: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
On Thu, Sep 03, 2015 at 09:38:16PM +0300, Adrian Hunter wrote:
SNIP
Quoted 11 lines
Quoted 4 lines
perf_evlist__propagate_mapsis called fromperf_evlist__create_maps, so if evsel is added later it will not be affected, perhaps we need something like belowYes but it would be nice to have a single function to do it that knows the rules for when the
evsel->cpusshould be retained. Say:
int__perf_evlist__propagate_maps(struct perf_evlist *evlist, struct perf_evsel *evsel)then call it fromperf_evlist__propagate_maps(),perf_evlist__add()andperf_evlist__splice_list_tail(). Probably need to add a member toperf_evlistto know when theevlist->cpusshould replace theevsel->cpus
ook,
jirka
Re: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
In reply to Jiri OlsaEm Thu, Sep 03, 2015 at 08:19:39PM +0200, Jiri Olsa escreveu:
Quoted 21 lines · Arnaldo Carvalho de Melo
Quoted 8 lines
Em Thu, Sep 03, 2015 at 07:23:43PM +0300, Adrian Hunter escreveu:
Quoted 6 lines · Arnaldo Carvalho de Melo
Quoted 5 lines
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 1 line · tip-bot for Kan Liang
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
Quoted 5 lines
Quoted 2 lines
There is "Test tracking with
sched_switch" but you need to expose it to the same issue i.e.Sure, Kan and Jiri were talking about the need to go doing these changes, Jiri? Kan?
perf_evlist__propagate_mapsis called fromperf_evlist__create_maps, so if evsel is added later it will not be affected, perhaps we need something like below:+++
b/tools/perf/util/evlist.c
tools/perf/util/evlist.c void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry) · line 138
Quoted 6 lines
tools/perf/util/evlist.c +3
-0@@ -133,6 +133,9 @@ void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)if (!evlist->nr_entries++)perf_evlist__set_id_pos (evlist); + + entry->cpus =cpu_map__get (evlist->cpus); + entry->threads =thread_map__get (evlist->threads);
You can't simply do that, we need to do it only if those fields are not already set.
- Arnaldo
Re: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Quoted 23 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 08:19:39PM +0200, Jiri Olsa escreveu:
Quoted 22 lines · Arnaldo Carvalho de Melo
Quoted 8 lines
Em Thu, Sep 03, 2015 at 07:23:43PM +0300, Adrian Hunter escreveu:
Quoted 6 lines · Arnaldo Carvalho de Melo
Quoted 5 lines
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 1 line · tip-bot for Kan Liang
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
Quoted 5 lines
Quoted 2 lines
There is "Test tracking with
sched_switch" but you need to expose it to the same issue i.e.Sure, Kan and Jiri were talking about the need to go doing these changes, Jiri? Kan?
perf_evlist__propagate_mapsis called fromperf_evlist__create_maps, so if evsel is added later it will not be affected, perhaps we need something like below:+++
b/tools/perf/util/evlist.c(quoted diff) +0
-0@@ -133,6 +133,9 @@ void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
Quoted 8 lines
tools/perf/util/evlist.c void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry) · line 138
Quoted 5 lines
tools/perf/util/evlist.c +3
-0if (!evlist->nr_entries++)perf_evlist__set_id_pos (evlist); + + entry->cpus =cpu_map__get (evlist->cpus); + entry->threads =thread_map__get (evlist->threads);You can't simply do that, we need to do it only if those fields are not already set.
argh, right.. forgot about cpus setup.. anyway, Adrian already wanted to do single function for this and the propagate function
jirka
Adrian Hunter
reviewerRe: [tip:perf/urgent] perf evlist: Open event on evsel cpus and threads
Quoted 33 lines · Jiri Olsa
Quoted 23 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 08:19:39PM +0200, Jiri Olsa escreveu:
Quoted 22 lines · Arnaldo Carvalho de Melo
Quoted 8 lines
Em Thu, Sep 03, 2015 at 07:23:43PM +0300, Adrian Hunter escreveu:
Quoted 6 lines · Arnaldo Carvalho de Melo
Quoted 5 lines
Em Thu, Sep 03, 2015 at 04:34:24PM +0300, Adrian Hunter escreveu:
Quoted 1 line · tip-bot for Kan Liang
Something we could catch in a 'test' entry? Even if that required Intel PT hardware that would be something important to have, all this stuff is growing in complexity, we need those tests...
Quoted 5 lines
Quoted 2 lines
There is "Test tracking with
sched_switch" but you need to expose it to the same issue i.e.Sure, Kan and Jiri were talking about the need to go doing these changes, Jiri? Kan?
perf_evlist__propagate_mapsis called fromperf_evlist__create_maps, so if evsel is added later it will not be affected, perhaps we need something like below:+++
b/tools/perf/util/evlist.c(quoted diff) +0
-0@@ -133,6 +133,9 @@ void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)Quoted 8 lines
tools/perf/util/evlist.c void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry) · line 138
Quoted 5 lines
tools/perf/util/evlist.c +3
-0if (!evlist->nr_entries++)perf_evlist__set_id_pos (evlist); + + entry->cpus =cpu_map__get (evlist->cpus); + entry->threads =thread_map__get (evlist->threads);You can't simply do that, we need to do it only if those fields are not already set.
Quoted 2 lines
argh, right.. forgot about cpus setup.. anyway, Adrian already wanted to do single function for this and the propagate function
yeah, I started working on it this morning.
Adrian Hunter
reviewerPATCH 1/1 perf tools: Fix gaps propagating maps
A perf_evsel is a selected event containing the perf_event_attr that is passed to perf_event_open(). A perf_evlist is a collection of perf_evsel's. A perf_evlist also has lists of cpus and threads (pids) on which to open the event. These lists are called 'maps' and this patch is about how those 'maps' are propagated from the perf_evlist to the perf_evsels.
Originally perf_evsels did not have their own thread 'maps', and their cpu 'maps' were only used for checking. Then threads were added by:
578e91ec04d0 ("perf evlist: Propagate thread maps through the evlist")
And then 'perf record' was changed to open events using the 'maps' from perf_evsel not perf_evlist anymore, by:
d988d5ee6478 ("perf evlist: Open event on evsel cpus and threads")
That exposed situations where the 'maps' were not getting propagated correctly, such as when perf_evsel are added to the perf_evlist later. This patch ensures 'maps' get propagated in that case, and also if 'maps' are subsequently changed or set to NULL by perf_evlist__set_maps() and perf_evlist__create_syswide_maps().
--- tools/perf/util/evlist.c | 129 ++++++++++++++++++++++++++++++++--------------- tools/perf/util/evlist.h | 1 + 2 files changed, 88 insertions(+), 42 deletions(-)
tools/perf/util/evlist.c +87 -42
@@ -124,8 +124,49 @@ void perf_evlist__delete(struct perf_evlist *evlist)free (evlist); } +static void __perf_evlist__propagate_maps (struct perf_evlist *evlist, +struct perf_evsel *evsel, +bool propagate) +{ +if (propagate) { +/* +* We already have cpus for evsel (via PMU sysfs) so +* keep it, if there's no target cpu list defined. +*/ +if ((!evsel->cpus || evlist->has_user_cpus) && + evsel->cpus != evlist->cpus) { +cpu_map__put (evsel->cpus); + evsel->cpus =cpu_map__get (evlist->cpus); + } + +if (!evsel->threads) + evsel->threads =thread_map__get (evlist->threads); + }else { +if (evsel->cpus == evlist->cpus) { +cpu_map__put (evsel->cpus); + evsel->cpus =NULL ; + } + +if (evsel->threads == evlist->threads) { +thread_map__put (evsel->threads); + evsel->threads =NULL ; + } + } +} + +static void perf_evlist__propagate_maps (struct perf_evlist *evlist, +bool propagate) +{ +struct perf_evsel *evsel; + +evlist__for_each (evlist, evsel) +__perf_evlist__propagate_maps (evlist, evsel, propagate); +} +void perf_evlist__add (struct perf_evlist *evlist,struct perf_evsel *entry) { +__perf_evlist__propagate_maps (evlist, entry, true); + entry->evlist = evlist;list_add_tail (&entry->node, &evlist->entries); entry->idx = evlist->nr_entries;
@@ -140,6 +181,10 @@ void perf_evlist__splice_list_tail(struct perf_evlist *evlist, int nr_entries) {bool set_id_pos = !evlist->nr_entries; +struct perf_evsel *evsel; + +__evlist__for_each (list, evsel) +__perf_evlist__propagate_maps (evlist, evsel, true);list_splice_tail (list, &evlist->entries); evlist->nr_entries += nr_entries;
@@ -1103,34 +1148,11 @@ int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); } -static int perf_evlist__propagate_maps (struct perf_evlist *evlist, -bool has_user_cpus) -{ -struct perf_evsel *evsel; - -evlist__for_each (evlist, evsel) { -/* -* We already have cpus for evsel (via PMU sysfs) so -* keep it, if there's no target cpu list defined. -*/ -if (evsel->cpus && has_user_cpus) -cpu_map__put (evsel->cpus); - -if (!evsel->cpus || has_user_cpus) - evsel->cpus =cpu_map__get (evlist->cpus); - - evsel->threads =thread_map__get (evlist->threads); - -if ((evlist->cpus && !evsel->cpus) || - (evlist->threads && !evsel->threads)) -return -ENOMEM ; - } - -return 0 ; -} -int perf_evlist__create_maps (struct perf_evlist *evlist,struct target *target) { +if (evlist->threads || evlist->cpus) +return -1 ; + evlist->threads =thread_map__new_str (target->pid, target->tid, target->uid);
@@ -1145,7 +1167,11 @@ int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)if (evlist->cpus ==NULL ) goto out_delete_threads; -return perf_evlist__propagate_maps (evlist, !!target->cpu_list); + evlist->has_user_cpus = !!target->cpu_list; + +perf_evlist__propagate_maps (evlist, true); + +return 0 ; out_delete_threads:thread_map__put (evlist->threads);
@@ -1157,17 +1183,32 @@ int perf_evlist__set_maps(struct perf_evlist *evlist,struct cpu_map *cpus, struct thread_map *threads) { -if (evlist->cpus) -cpu_map__put (evlist->cpus); +/* +* First 'un-propagate' the current maps which allows the propagation to +* work correctly even when changing the maps or setting them to NULL. +*/ +perf_evlist__propagate_maps (evlist ,false ); - evlist->cpus = cpus; +/* +* Allow for the possibility that one or another of the maps isn't being +* changed i.e. don't put it. Note we are assuming the maps that are +* being applied are brand new and evlist is taking ownership of the +* original reference count of 1. If that is not the case it is up to +* the caller to increase the reference count. +*/ +if (cpus != evlist->cpus) { +cpu_map__put (evlist->cpus); + evlist->cpus = cpus; + } -if (evlist->threads) +if (threads != evlist->threads) {thread_map__put (evlist->threads); + evlist->threads = threads; + } - evlist->threads = threads; +perf_evlist__propagate_maps (evlist, true); -return perf_evlist__propagate_maps (evlist, false); +return 0 ; }int perf_evlist__apply_filters (struct perf_evlist *evlist,struct perf_evsel **err_evsel)
@@ -1387,6 +1428,8 @@ void perf_evlist__close(struct perf_evlist *evlist)static int perf_evlist__create_syswide_maps (struct perf_evlist *evlist) { +struct cpu_map *cpus; +struct thread_map *threads;int err = -ENOMEM ;/*
@@ -1398,20 +1441,22 @@ static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)* error, and we may not want to do that fallback to a * default cpu identity map :-\ */ - evlist->cpus =cpu_map__new (NULL ); -if (evlist->cpus ==NULL ) + cpus =cpu_map__new (NULL ); +if (!cpus) goto out; - evlist->threads =thread_map__new_dummy (); -if (evlist->threads ==NULL ) - goto out_free_cpus; + threads =thread_map__new_dummy (); +if (!threads) + goto out_put; - err =0 ; + err =perf_evlist__set_maps (evlist, cpus, threads); +if (err) + goto out_put; out:return err; -out_free_cpus: -cpu_map__put (evlist->cpus); - evlist->cpus =NULL ; +out_put: +cpu_map__put (cpus); +thread_map__put (threads); goto out; }
tools/perf/util/evlist.h +1 -0
@@ -42,6 +42,7 @@ struct perf_evlist {int nr_mmaps;bool overwrite;bool enabled; +bool has_user_cpus;size_t mmap_len;int id_pos;int is_pos;
Signature · 1 line
1.9.1
Re: [PATCH] perf tools: Fix gaps propagating maps
Em Fri, Sep 04, 2015 at 03:15:54PM +0300, Adrian Hunter escreveu:
Quoted 6 lines
A
perf_evselis a selected event containing theperf_event_attrthat is passed toperf_event_open(). Aperf_evlistis a collection of perf_evsel's. Aperf_evlistalso has lists of cpus and threads (pids) on which to open the event. These lists are called 'maps' and this patch is about how those 'maps' are propagated from theQuoted 1 line
perf_evlistto theperf_evsels.
Can't this be broken up in multiple patches, for instance this:
int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
{
+ if (evlist->threads || evlist->cpus)
+ return -1;
+
Looks like a fix that could be separated. Also FOO__propagate(.., false) to do the opposite of propagate seems confusing, how about FOO__unpropagate() if that verb exists? :-)
Also, when unpropagating, you do:
if (evsel->cpus == evlist->cpus) {
cpu_map__put(evsel->cpus);
evsel->cpus = NULL;
}What if the PMU code _set_ it to the same cpus as in evlist->cpus, but now we're unpropagating to set to another CPU, in this case we will be changing the PMU setting with a new one. I.e. when a PMU sets it it should be sticky, no?
I.e. we would have to know, in the evsel, if evsel->cpus was set by the PMU or any other future entity expecting this behaviour, so that we don't touch it, i.e. testing (evsel->cpus != evlist->cpus) when unpropagating doesn't seem to cut, right?
- Arnaldo
Quoted 233 lines
Originally
perf_evselsdid not have their own thread 'maps', and their cpu 'maps' were only used for checking. Then threads were added by:
578e91ec04d0("perf evlist: Propagate thread maps through the evlist")And then 'perf record' was changed to open events using the 'maps' from
perf_evselnotperf_evlistanymore, by:
d988d5ee6478("perf evlist: Open event on evsel cpus and threads")That exposed situations where the 'maps' were not getting propagated correctly, such as when
perf_evselare added to theperf_evlistlater. This patch ensures 'maps' get propagated in that case, and also if 'maps' are subsequently changed or set to NULL byperf_evlist__set_maps()andperf_evlist__create_syswide_maps().Signed-off-by Adrian Hunter <[email protected]>--- tools/perf/util/evlist.c | 129 ++++++++++++++++++++++++++++++++--------------- tools/perf/util/evlist.h | 1 + 2 files changed, 88 insertions(+), 42 deletions(-)tools/perf/util/evlist.c +87
-42@@ -124,8 +124,49 @@ void perf_evlist__delete(struct perf_evlist *evlist)free (evlist); } +static void __perf_evlist__propagate_maps (struct perf_evlist *evlist, +struct perf_evsel *evsel, +bool propagate) +{ +if (propagate) { +/* +* We already have cpus for evsel (via PMU sysfs) so +* keep it, if there's no target cpu list defined. +*/ +if ((!evsel->cpus || evlist->has_user_cpus) && + evsel->cpus != evlist->cpus) { +cpu_map__put (evsel->cpus); + evsel->cpus =cpu_map__get (evlist->cpus); + } + +if (!evsel->threads) + evsel->threads =thread_map__get (evlist->threads); + }else { +if (evsel->cpus == evlist->cpus) { +cpu_map__put (evsel->cpus); + evsel->cpus =NULL ; + } + +if (evsel->threads == evlist->threads) { +thread_map__put (evsel->threads); + evsel->threads =NULL ; + } + } +} + +static void perf_evlist__propagate_maps (struct perf_evlist *evlist, +bool propagate) +{ +struct perf_evsel *evsel; + +evlist__for_each (evlist, evsel) +__perf_evlist__propagate_maps (evlist, evsel, propagate); +} +void perf_evlist__add (struct perf_evlist *evlist,struct perf_evsel *entry) { +__perf_evlist__propagate_maps (evlist, entry, true); + entry->evlist = evlist;list_add_tail (&entry->node, &evlist->entries); entry->idx = evlist->nr_entries;@@ -140,6 +181,10 @@ void perf_evlist__splice_list_tail(struct perf_evlist *evlist, int nr_entries) {bool set_id_pos = !evlist->nr_entries; +struct perf_evsel *evsel; + +__evlist__for_each (list, evsel) +__perf_evlist__propagate_maps (evlist, evsel, true);list_splice_tail (list, &evlist->entries); evlist->nr_entries += nr_entries;@@ -1103,34 +1148,11 @@ int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); } -static int perf_evlist__propagate_maps (struct perf_evlist *evlist, -bool has_user_cpus) -{ -struct perf_evsel *evsel; - -evlist__for_each (evlist, evsel) { -/* -* We already have cpus for evsel (via PMU sysfs) so -* keep it, if there's no target cpu list defined. -*/ -if (evsel->cpus && has_user_cpus) -cpu_map__put (evsel->cpus); - -if (!evsel->cpus || has_user_cpus) - evsel->cpus =cpu_map__get (evlist->cpus); - - evsel->threads =thread_map__get (evlist->threads); - -if ((evlist->cpus && !evsel->cpus) || - (evlist->threads && !evsel->threads)) -return -ENOMEM ; - } - -return 0 ; -} -int perf_evlist__create_maps (struct perf_evlist *evlist,struct target *target) { +if (evlist->threads || evlist->cpus) +return -1 ; + evlist->threads =thread_map__new_str (target->pid, target->tid, target->uid);@@ -1145,7 +1167,11 @@ int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)if (evlist->cpus ==NULL ) goto out_delete_threads; -return perf_evlist__propagate_maps (evlist, !!target->cpu_list); + evlist->has_user_cpus = !!target->cpu_list; + +perf_evlist__propagate_maps (evlist, true); + +return 0 ; out_delete_threads:thread_map__put (evlist->threads);@@ -1157,17 +1183,32 @@ int perf_evlist__set_maps(struct perf_evlist *evlist,struct cpu_map *cpus, struct thread_map *threads) { -if (evlist->cpus) -cpu_map__put (evlist->cpus); +/* +* First 'un-propagate' the current maps which allows the propagation to +* work correctly even when changing the maps or setting them to NULL. +*/ +perf_evlist__propagate_maps (evlist ,false ); - evlist->cpus = cpus; +/* +* Allow for the possibility that one or another of the maps isn't being +* changed i.e. don't put it. Note we are assuming the maps that are +* being applied are brand new and evlist is taking ownership of the +* original reference count of 1. If that is not the case it is up to +* the caller to increase the reference count. +*/ +if (cpus != evlist->cpus) { +cpu_map__put (evlist->cpus); + evlist->cpus = cpus; + } -if (evlist->threads) +if (threads != evlist->threads) {thread_map__put (evlist->threads); + evlist->threads = threads; + } - evlist->threads = threads; +perf_evlist__propagate_maps (evlist, true); -return perf_evlist__propagate_maps (evlist, false); +return 0 ; }int perf_evlist__apply_filters (struct perf_evlist *evlist,struct perf_evsel **err_evsel)@@ -1387,6 +1428,8 @@ void perf_evlist__close(struct perf_evlist *evlist)static int perf_evlist__create_syswide_maps (struct perf_evlist *evlist) { +struct cpu_map *cpus; +struct thread_map *threads;int err = -ENOMEM ;/* @@ -1398,20 +1441,22 @@ static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)* error, and we may not want to do that fallback to a * default cpu identity map :-\ */ - evlist->cpus =cpu_map__new (NULL ); -if (evlist->cpus ==NULL ) + cpus =cpu_map__new (NULL ); +if (!cpus) goto out; - evlist->threads =thread_map__new_dummy (); -if (evlist->threads ==NULL ) - goto out_free_cpus; + threads =thread_map__new_dummy (); +if (!threads) + goto out_put; - err =0 ; + err =perf_evlist__set_maps (evlist, cpus, threads); +if (err) + goto out_put; out:return err; -out_free_cpus: -cpu_map__put (evlist->cpus); - evlist->cpus =NULL ; +out_put: +cpu_map__put (cpus); +thread_map__put (threads); goto out; }tools/perf/util/evlist.h +1
-0@@ -42,6 +42,7 @@ struct perf_evlist {int nr_mmaps;bool overwrite;bool enabled; +bool has_user_cpus;size_t mmap_len;int id_pos;int is_pos;Signature · 1 line
1.9.1
Adrian Hunter
reviewerRe: [PATCH] perf tools: Fix gaps propagating maps
Quoted 9 lines · Arnaldo Carvalho de Melo
Em Fri, Sep 04, 2015 at 03:15:54PM +0300, Adrian Hunter escreveu:
Quoted 6 lines
A
perf_evselis a selected event containing theperf_event_attrthat is passed toperf_event_open(). Aperf_evlistis a collection of perf_evsel's. Aperf_evlistalso has lists of cpus and threads (pids) on which to open the event. These lists are called 'maps' and this patch is about how those 'maps' are propagated from theQuoted 1 line
perf_evlistto theperf_evsels.Can't this be broken up in multiple patches, for instance this:
Ok, might not be until next week though.
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1155
Quoted 7 lines
int
perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
{tools/perf/util/evlist.c +3
-0+if (evlist->threads || evlist->cpus) +return -1 ; +
Or you could just drop that chunk.
Quoted 4 lines
Looks like a fix that could be separated. Also
FOO__propagate(.., false)to do the opposite of propagate seems confusing, how aboutFOO__unpropagate()if that verb exists? :-)
Ok
Quoted 9 lines
Also, when unpropagating, you do:
if (evsel->cpus == evlist->cpus) { cpu_map__put(evsel->cpus); evsel->cpus = NULL; }
Quoted 4 lines
What if the PMU code _set_ it to the same cpus as in
evlist->cpus, but now we're unpropagating to set to another CPU, in this case we will be changing the PMU setting with a new one. I.e. when a PMU sets it it should be sticky, no?
We are comparing the pointer, so that won't happen unless the PMU actually does evsel->cpus = evlist->cpus which seems unlikely.
Quoted 5 lines
I.e. we would have to know, in the evsel, if
evsel->cpuswas set by the PMU or any other future entity expecting this behaviour, so that we don't touch it, i.e. testing (evsel->cpus!=evlist->cpus) when unpropagating doesn't seem to cut, right?
I think the pointer comparison covers that. i.e. the pointers won't be the same even if the cpus are.
Re: [PATCH] perf tools: Fix gaps propagating maps
Em Fri, Sep 04, 2015 at 04:42:30PM +0300, Adrian Hunter escreveu:
Quoted 7 lines · Arnaldo Carvalho de Melo
Quoted 6 lines
Em Fri, Sep 04, 2015 at 03:15:54PM +0300, Adrian Hunter escreveu:
Quoted 5 lines
A
perf_evselis a selected event containing theperf_event_attrthat is passed toperf_event_open(). Aperf_evlistis a collection of perf_evsel's. Aperf_evlistalso has lists of cpus and threads (pids) on which to open the event. These lists are called 'maps' and this patch is about how those 'maps' are propagated from the
Quoted 5 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
perf_evlistto theperf_evsels.Can't this be broken up in multiple patches, for instance this:
Ok, might not be until next week though.
Take your time, it seems that so far the only problem is with Intel PT, with adding that sched_switch after the propagation was done, no?
And with kernel with PERF_RECORD_SWITCH, that I am pushing that other patch from you, to the support using it replacing sched:sched_Switch in the tooling side reduces the impact of this problem, right?
Quoted 30 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1155
Quoted 6 lines
int
perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
{tools/perf/util/evlist.c +3
-0+if (evlist->threads || evlist->cpus) +return -1 ; +Or you could just drop that chunk.
Quoted 3 lines
Looks like a fix that could be separated. Also
FOO__propagate(.., false)to do the opposite of propagate seems confusing, how aboutFOO__unpropagate()if that verb exists? :-)Ok
Quoted 11 lines
Also, when unpropagating, you do:
if (evsel->cpus == evlist->cpus) { cpu_map__put(evsel->cpus); evsel->cpus = NULL; }What if the PMU code _set_ it to the same cpus as in
evlist->cpus, but now we're unpropagating to set to another CPU, in this case we will be changing the PMU setting with a new one. I.e. when a PMU sets it it should be sticky, no?We are comparing the pointer, so that won't happen unless the PMU actually does
evsel->cpus=evlist->cpuswhich seems unlikely.
Quoted 7 lines
Quoted 4 lines
I.e. we would have to know, in the evsel, if
evsel->cpuswas set by the PMU or any other future entity expecting this behaviour, so that we don't touch it, i.e. testing (evsel->cpus!=evlist->cpus) when unpropagating doesn't seem to cut, right?I think the pointer comparison covers that. i.e. the pointers won't be the same even if the cpus are.
It makes it more unlikely, yes.
I think that to be safe we should have a evsel->sticky_{cpu,threads}, well, at least the evsel->sticky_cpus seems needed due to the PMU usecase.
- Arnaldo