perf evlist: Allow setting arbitrary leader
Ian Rogers
reviewerPATCH 1/2 perf evlist: Allow setting arbitrary leader
The leader of a group is the first, but allow it to be an arbitrary list member so that for Intel topdown events slots may always be the group leader.
--- tools/lib/perf/evlist.c | 15 +++++++++------ tools/lib/perf/include/internal/evlist.h | 2 +- tools/perf/util/parse-events.c | 2 +- 3 files changed, 11 insertions(+), 8 deletions(-)
tools/lib/perf/evlist.c +9 -6
@@ -643,14 +643,14 @@ perf_evlist__next_mmap(struct perf_evlist *evlist, struct perf_mmap *map,return overwrite ? evlist->mmap_ovw_first : evlist->mmap_first; } -void __perf_evlist__set_leader (struct list_head *list) +void __perf_evlist__set_leader (struct list_head *list,struct perf_evsel *leader) { -struct perf_evsel *evsel, *leader; +struct perf_evsel *first, *last, *evsel; - leader = list_entry(list->next, struct perf_evsel, node); - evsel = list_entry(list->prev, struct perf_evsel, node); + first = list_entry(list->next, struct perf_evsel, node); + last = list_entry(list->prev, struct perf_evsel, node); - leader->nr_members = evsel->idx - leader->idx +1 ; + leader->nr_members = last->idx - first->idx +1 ;__perf_evlist__for_each_entry (list, evsel) evsel->leader = leader;
@@ -659,7 +659,10 @@ void __perf_evlist__set_leader(struct list_head *list)void perf_evlist__set_leader (struct perf_evlist *evlist) {if (evlist->nr_entries) { +struct perf_evsel *first =list_entry (evlist->entries.next, + struct perf_evsel, node); + evlist->nr_groups = evlist->nr_entries >1 ?1 :0 ; -__perf_evlist__set_leader (&evlist->entries); +__perf_evlist__set_leader (&evlist->entries, first); } }
tools/lib/perf/include/internal/evlist.h +1 -1
@@ -127,5 +127,5 @@ int perf_evlist__id_add_fd(struct perf_evlist *evlist,void perf_evlist__reset_id_hash (struct perf_evlist *evlist); -void __perf_evlist__set_leader (struct list_head *list); +void __perf_evlist__set_leader (struct list_head *list,struct perf_evsel *leader);#endif /* __LIBPERF_INTERNAL_EVLIST_H */
tools/perf/util/parse-events.c +1 -1
@@ -1834,8 +1834,8 @@ void parse_events__set_leader(char *name, struct list_head *list,if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state))return ; -__perf_evlist__set_leader (list); leader = list_entry(list->next, struct evsel, core.node); +__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; }
Signature · 1 line
2.34.0.rc1.387.gb447b232ab-goog
Ian Rogers
reviewerPATCH 2/2 perf parse-events: Architecture specific leader override
Currently topdown events must appear after a slots event:
$ perf stat -e '{slots,topdown-fe-bound}' /bin/true Performance counter stats for '/bin/true':
3,183,090 slots
986,133 topdown-fe-boundReversing the events yields:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (topdown-fe-bound).For metrics the order of events is determined by iterating over a hashmap, and so slots isn't guaranteed to be first which can yield this error.
Change the set_leader in parse-events, called when a group is closed, so that rather than always making the first event the leader, if the slots event exists then it is made the leader. It is then moved to the head of the evlist otherwise it won't be opened in the correct order.
The result is:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true Performance counter stats for '/bin/true':
3,274,795 slots
1,001,702 topdown-fe-boundA problem with this approach is the slots event is identified by name, names can be overwritten like 'cpu/slots,name=foo/' and this causes the leader change to fail.
--- tools/perf/arch/x86/util/evlist.c | 17 +++++++++++++++++ tools/perf/util/evlist.h | 1 + tools/perf/util/parse-events.c | 16 +++++++++++----- tools/perf/util/parse-events.h | 4 ++-- tools/perf/util/parse-events.y | 12 ++++++++---- 5 files changed, 39 insertions(+), 11 deletions(-)
tools/perf/arch/x86/util/evlist.c +17 -0
@@ -17,3 +17,20 @@ int arch_evlist__add_default_attrs(struct evlist *evlist)else return parse_events (evlist,TOPDOWN_L1_EVENTS ,NULL ); } + +struct evsel *arch_evlist__leader (struct list_head *list) +{ +struct evsel *evsel, *first; + + first = list_entry(list->next, struct evsel, core.node); + +if (!pmu_have_event ("cpu" ,"slots" )) +return first; + +__evlist__for_each_entry (list, evsel) { +if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && + evsel->name &&strstr (evsel->name,"slots" )) +return evsel; + } +return first; +}
tools/perf/util/evlist.h +1 -0
@@ -110,6 +110,7 @@ int __evlist__add_default_attrs(struct evlist *evlist, __evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))int arch_evlist__add_default_attrs (struct evlist *evlist); +struct evsel *arch_evlist__leader (struct list_head *list);int evlist__add_dummy (struct evlist *evlist);
tools/perf/util/parse-events.c +11 -5
@@ -1821,22 +1821,28 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,return ret; } -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state) +__weak struct evsel *arch_evlist__leader (struct list_head *list) +{ + return list_entry(list->next, struct evsel, core.node); +} + +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state) {struct evsel *leader;if (list_empty (list)) {WARN_ONCE (true,"WARNING: failed to set leader: empty list" ); -return ; +return NULL ; }if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state)) -return ; +return NULL ; - leader = list_entry(list->next, struct evsel, core.node); + leader =arch_evlist__leader (list);__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; +return &leader->core.node; }/* list_event is assumed to point to malloc'ed memory */
tools/perf/util/parse-events.h +2 -2
@@ -211,8 +211,8 @@ int parse_events_copy_term_list(struct list_head *old,enum perf_pmu_event_symbol_type perf_pmu__parse_check (const char *name); -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state); +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state);void parse_events_update_lists (struct list_head *list_event,struct list_head *list_all); void parse_events_evlist_error(struct parse_events_state *parse_state,
tools/perf/util/parse-events.y +8 -4
@@ -199,20 +199,24 @@ group_def group_def: PE_NAME '{' events '}' { - struct list_head *list = $3; + struct list_head *new_head, *list = $3; inc_group_count(list, _parse_state); - parse_events__set_leader($1, list, _parse_state); + new_head = parse_events__set_leader($1, list, _parse_state); + if (new_head) + list_move(new_head, list); free($1); $$ = list; } | '{' events '}' { - struct list_head *list = $2; + struct list_head *new_head, *list = $2; inc_group_count(list, _parse_state); - parse_events__set_leader(NULL, list, _parse_state); + new_head = parse_events__set_leader(NULL, list, _parse_state); + if (new_head) + list_move(new_head, list); $$ = list; }
Signature · 1 line
2.34.0.rc1.387.gb447b232ab-goog
Re: [PATCH 2/2] perf parse-events: Architecture specific leader override
Quoted 6 lines · Ian Rogers
Currently topdown events must appear after a slots event:
$ perf stat -e '{slots,topdown-fe-bound}' /bin/truePerformance counter stats for '
/bin/true':
Quoted 8 lines
3,183,090 slots 986,133 topdown-fe-boundReversing the events yields:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true Error: The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (topdown-fe-bound).
why is this actually failing?
Quoted 14 lines
For metrics the order of events is determined by iterating over a hashmap, and so slots isn't guaranteed to be first which can yield this error.
Change the
set_leaderin parse-events, called when a group is closed, so that rather than always making the first event the leader, if the slots event exists then it is made the leader. It is then moved to the head of the evlist otherwise it won't be opened in the correct order.The result is:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true
Quoted 8 lines
Performance counter stats for '
/bin/true':3,274,795 slots 1,001,702 topdown-fe-boundA problem with this approach is the slots event is identified by name, names can be overwritten like 'cpu/slots,name=foo/' and this causes the leader change to fail.
would it be better then to move this shuffle to the metric code directly, and make sure it only spits 'good' order of events?
and keep "-e '{topdown-fe-bound,slots}'" to fail if user specifies that on the command line
Quoted 106 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 17 +++++++++++++++++ tools/perf/util/evlist.h | 1 + tools/perf/util/parse-events.c | 16 +++++++++++----- tools/perf/util/parse-events.h | 4 ++-- tools/perf/util/parse-events.y | 12 ++++++++---- 5 files changed, 39 insertions(+), 11 deletions(-)tools/perf/arch/x86/util/evlist.c +17
-0@@ -17,3 +17,20 @@ int arch_evlist__add_default_attrs(struct evlist *evlist)else return parse_events (evlist,TOPDOWN_L1_EVENTS ,NULL ); } + +struct evsel *arch_evlist__leader (struct list_head *list) +{ +struct evsel *evsel, *first; + + first = list_entry(list->next, struct evsel, core.node); + +if (!pmu_have_event ("cpu" ,"slots" )) +return first; + +__evlist__for_each_entry (list, evsel) { +if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && + evsel->name &&strstr (evsel->name,"slots" )) +return evsel; + } +return first; +}tools/perf/util/evlist.h +1
-0@@ -110,6 +110,7 @@ int __evlist__add_default_attrs(struct evlist *evlist, __evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))int arch_evlist__add_default_attrs (struct evlist *evlist); +struct evsel *arch_evlist__leader (struct list_head *list);int evlist__add_dummy (struct evlist *evlist);tools/perf/util/parse-events.c +11
-5@@ -1821,22 +1821,28 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,return ret; } -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state) +__weak struct evsel *arch_evlist__leader (struct list_head *list) +{ + return list_entry(list->next, struct evsel, core.node); +} + +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state) {struct evsel *leader;if (list_empty (list)) {WARN_ONCE (true,"WARNING: failed to set leader: empty list" ); -return ; +return NULL ; }if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state)) -return ; +return NULL ; - leader = list_entry(list->next, struct evsel, core.node); + leader =arch_evlist__leader (list);__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; +return &leader->core.node; }/* list_event is assumed to point to malloc'ed memory */ tools/perf/util/parse-events.h +2
-2@@ -211,8 +211,8 @@ int parse_events_copy_term_list(struct list_head *old,enum perf_pmu_event_symbol_type perf_pmu__parse_check (const char *name); -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state); +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state);void parse_events_update_lists (struct list_head *list_event,struct list_head *list_all); void parse_events_evlist_error(struct parse_events_state *parse_state,tools/perf/util/parse-events.y +0
-0@@ -199,20 +199,24 @@ group_def group_def:
tools/perf/util/parse-events.y group_def · line 207
Quoted 10 lines
tools/perf/util/parse-events.y +4
-2PE_NAME '{' events '}' { - struct list_head *list = $3; + struct list_head *new_head, *list = $3; inc_group_count(list, _parse_state); - parse_events__set_leader($1, list, _parse_state); + new_head = parse_events__set_leader($1, list, _parse_state); + if (new_head) + list_move(new_head, list);
why not put these last 2 lines to parse_events__set_leader as well?
Quoted 4 lines
tools/perf/util/parse-events.y +0
-0free($1); $$ = list; } |
tools/perf/util/parse-events.y group_def · line 219
Quoted 10 lines
tools/perf/util/parse-events.y +4
-2'{' events '}' { - struct list_head *list = $2; + struct list_head *new_head, *list = $2; inc_group_count(list, _parse_state); - parse_events__set_leader(NULL, list, _parse_state); + new_head = parse_events__set_leader(NULL, list, _parse_state); + if (new_head) + list_move(new_head, list);
same
thanks,
jirka
Ian Rogers
reviewerRe: [PATCH 2/2] perf parse-events: Architecture specific leader override
Quoted 42 lines · Jiri Olsa
Quoted 6 lines · Ian Rogers
Currently topdown events must appear after a slots event:
$ perf stat -e '{slots,topdown-fe-bound}' /bin/truePerformance counter stats for '
/bin/true':Quoted 8 lines
3,183,090 slots 986,133 topdown-fe-boundReversing the events yields:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true Error: The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (topdown-fe-bound).why is this actually failing?
Quoted 14 lines
For metrics the order of events is determined by iterating over a hashmap, and so slots isn't guaranteed to be first which can yield this error.
Change the
set_leaderin parse-events, called when a group is closed, so that rather than always making the first event the leader, if the slots event exists then it is made the leader. It is then moved to the head of the evlist otherwise it won't be opened in the correct order.The result is:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/trueQuoted 8 lines
Performance counter stats for '
/bin/true':3,274,795 slots 1,001,702 topdown-fe-boundA problem with this approach is the slots event is identified by name, names can be overwritten like 'cpu/slots,name=foo/' and this causes the leader change to fail.
Quoted 5 lines
would it be better then to move this shuffle to the metric code directly, and make sure it only spits 'good' order of events?
and keep "-e '{topdown-fe-bound,slots}'" to fail if user specifies that on the command line
It's an alternative to do that, but the people I spoke to preferred having parse-events do this. I'm not sure what Andi's opinion is. There is a general frustration about how brittle the slots event is, and so this does make it less brittle.
Quoted 111 lines
Quoted 111 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 17 +++++++++++++++++ tools/perf/util/evlist.h | 1 + tools/perf/util/parse-events.c | 16 +++++++++++----- tools/perf/util/parse-events.h | 4 ++-- tools/perf/util/parse-events.y | 12 ++++++++---- 5 files changed, 39 insertions(+), 11 deletions(-)tools/perf/arch/x86/util/evlist.c +17
-0@@ -17,3 +17,20 @@ int arch_evlist__add_default_attrs(struct evlist *evlist)else return parse_events (evlist,TOPDOWN_L1_EVENTS ,NULL ); } + +struct evsel *arch_evlist__leader (struct list_head *list) +{ +struct evsel *evsel, *first; + + first = list_entry(list->next, struct evsel, core.node); + +if (!pmu_have_event ("cpu" ,"slots" )) +return first; + +__evlist__for_each_entry (list, evsel) { +if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && + evsel->name &&strstr (evsel->name,"slots" )) +return evsel; + } +return first; +}tools/perf/util/evlist.h +1
-0@@ -110,6 +110,7 @@ int __evlist__add_default_attrs(struct evlist *evlist, __evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))int arch_evlist__add_default_attrs (struct evlist *evlist); +struct evsel *arch_evlist__leader (struct list_head *list);int evlist__add_dummy (struct evlist *evlist);tools/perf/util/parse-events.c +11
-5@@ -1821,22 +1821,28 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,return ret; } -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state) +__weak struct evsel *arch_evlist__leader (struct list_head *list) +{ + return list_entry(list->next, struct evsel, core.node); +} + +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state) {struct evsel *leader;if (list_empty (list)) {WARN_ONCE (true,"WARNING: failed to set leader: empty list" ); -return ; +return NULL ; }if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state)) -return ; +return NULL ; - leader = list_entry(list->next, struct evsel, core.node); + leader =arch_evlist__leader (list);__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; +return &leader->core.node; }/* list_event is assumed to point to malloc'ed memory */ tools/perf/util/parse-events.h +2
-2@@ -211,8 +211,8 @@ int parse_events_copy_term_list(struct list_head *old,enum perf_pmu_event_symbol_type perf_pmu__parse_check (const char *name); -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state); +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state);void parse_events_update_lists (struct list_head *list_event,struct list_head *list_all); void parse_events_evlist_error(struct parse_events_state *parse_state,tools/perf/util/parse-events.y +1
-1@@ -199,20 +199,24 @@ group_def group_def: PE_NAME '{' events '}' { - struct list_head *list = $3; + struct list_head *new_head, *list = $3;
Quoted 7 lines
tools/perf/util/parse-events.y group_def · line 207
Quoted 5 lines
tools/perf/util/parse-events.y +3
-1inc_group_count(list, _parse_state); - parse_events__set_leader($1, list, _parse_state); + new_head = parse_events__set_leader($1, list, _parse_state); + if (new_head) + list_move(new_head, list);why not put these last 2 lines to
parse_events__set_leaderas well?
I can move that, but I'll try to motivate having it this way. The list logic in Linux I find troublesome. In parse-events.y we have a list head that is just a next, prev pointer and not full evsel in a list. So for {topdown-fe-bound,slots} in parse-events.y we have a list of:
head -next-> evsel("topdown-fe-bound") -next-> evsel("slots")
But in __perf_evlist__set_leader the list doesn't have the head:
evsel("topdown-fe-bound") -next-> evsel("slots")
The list_move is changing the list to:
head -next-> evsel("slots") -next-> evsel("topdown-fe-bound")
which isn't possible in __perf_evlist__set_leader as the head has gone. I felt having the list_move in parse_events__set_leader made it look like the leader is being placed second rather than at the head. So this was an attempt to make the code more intention revealing, but that could also be done by some comments, it's just in parse_events__set_leader we have the two styles of lists in play.
While I'm complaining, there are many brittle assumptions about evlist that it may be worth adding some more assertions for. There is already an assertion that the leader was opened before its siblings. There are assumptions that the idx given to an evsel are incremental and that for the leader 'idx - prev->idx' will yield the number of list elements. My concern is for libperf users, it doesn't seem unreasonable that they will want to manipulate the evlist but given the assumptions they will likely find bugs - such as idx being used to index an array and reordering breaking things. The metric code would be a customer of reordering were it to need to move slots, but when I experimented with this it easily broke things - which motivates doing it in parse-events or more likely for the metrics, when generating the parse-events string.
Thanks,
Ian
Quoted 20 lines
Quoted 4 lines
tools/perf/util/parse-events.y +0
-0free($1); $$ = list; } |tools/perf/util/parse-events.y group_def · line 219
Quoted 10 lines
tools/perf/util/parse-events.y +4
-2'{' events '}' { - struct list_head *list = $2; + struct list_head *new_head, *list = $2; inc_group_count(list, _parse_state); - parse_events__set_leader(NULL, list, _parse_state); + new_head = parse_events__set_leader(NULL, list, _parse_state); + if (new_head) + list_move(new_head, list);same
thanks,
jirka
Ian Rogers
reviewerRe: [PATCH 2/2] perf parse-events: Architecture specific leader override
Quoted 47 lines · Ian Rogers
Quoted 45 lines · Jiri Olsa
Quoted 6 lines · Ian Rogers
Currently topdown events must appear after a slots event:
$ perf stat -e '{slots,topdown-fe-bound}' /bin/truePerformance counter stats for '
/bin/true':Quoted 8 lines
3,183,090 slots 986,133 topdown-fe-boundReversing the events yields:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true Error: The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (topdown-fe-bound).why is this actually failing?
Quoted 14 lines
For metrics the order of events is determined by iterating over a hashmap, and so slots isn't guaranteed to be first which can yield this error.
Change the
set_leaderin parse-events, called when a group is closed, so that rather than always making the first event the leader, if the slots event exists then it is made the leader. It is then moved to the head of the evlist otherwise it won't be opened in the correct order.The result is:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/trueQuoted 8 lines
Performance counter stats for '
/bin/true':3,274,795 slots 1,001,702 topdown-fe-boundA problem with this approach is the slots event is identified by name, names can be overwritten like 'cpu/slots,name=foo/' and this causes the leader change to fail.
would it be better then to move this shuffle to the metric code directly, and make sure it only spits 'good' order of events?
Quoted 7 lines
Quoted 2 lines
and keep "-e '{topdown-fe-bound,slots}'" to fail if user specifies that on the command line
It's an alternative to do that, but the people I spoke to preferred having parse-events do this. I'm not sure what Andi's opinion is. There is a general frustration about how brittle the slots event is, and so this does make it less brittle.
A different problem this fixes is if you have a group like '{instructions,slots,topdown-fe-bound}' then slots still has to be the group leader even though it is appearing before the topdown event. So before the patch:
$ perf stat -e '{instructions,slots,topdown-fe-bound}' -a -- sleep 2Performance counter stats for 'system wide':
<not counted> instructions
<not counted> slots
<not supported> topdown-fe-bound2.006410898 seconds time elapsed
After:
$ perf stat -e '{instructions,slots,topdown-fe-bound}' -a -- sleep 2Performance counter stats for 'system wide':
2572632080 slots
362537420 instructions
742416906 topdown-fe-bound2.005875050 seconds time elapsed
Thanks,
Ian
Quoted 181 lines
Quoted 111 lines
Quoted 111 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 17 +++++++++++++++++ tools/perf/util/evlist.h | 1 + tools/perf/util/parse-events.c | 16 +++++++++++----- tools/perf/util/parse-events.h | 4 ++-- tools/perf/util/parse-events.y | 12 ++++++++---- 5 files changed, 39 insertions(+), 11 deletions(-)tools/perf/arch/x86/util/evlist.c +17
-0@@ -17,3 +17,20 @@ int arch_evlist__add_default_attrs(struct evlist *evlist)else return parse_events (evlist,TOPDOWN_L1_EVENTS ,NULL ); } + +struct evsel *arch_evlist__leader (struct list_head *list) +{ +struct evsel *evsel, *first; + + first = list_entry(list->next, struct evsel, core.node); + +if (!pmu_have_event ("cpu" ,"slots" )) +return first; + +__evlist__for_each_entry (list, evsel) { +if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && + evsel->name &&strstr (evsel->name,"slots" )) +return evsel; + } +return first; +}tools/perf/util/evlist.h +1
-0@@ -110,6 +110,7 @@ int __evlist__add_default_attrs(struct evlist *evlist, __evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))int arch_evlist__add_default_attrs (struct evlist *evlist); +struct evsel *arch_evlist__leader (struct list_head *list);int evlist__add_dummy (struct evlist *evlist);tools/perf/util/parse-events.c +11
-5@@ -1821,22 +1821,28 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,return ret; } -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state) +__weak struct evsel *arch_evlist__leader (struct list_head *list) +{ + return list_entry(list->next, struct evsel, core.node); +} + +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state) {struct evsel *leader;if (list_empty (list)) {WARN_ONCE (true,"WARNING: failed to set leader: empty list" ); -return ; +return NULL ; }if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state)) -return ; +return NULL ; - leader = list_entry(list->next, struct evsel, core.node); + leader =arch_evlist__leader (list);__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; +return &leader->core.node; }/* list_event is assumed to point to malloc'ed memory */ tools/perf/util/parse-events.h +2
-2@@ -211,8 +211,8 @@ int parse_events_copy_term_list(struct list_head *old,enum perf_pmu_event_symbol_type perf_pmu__parse_check (const char *name); -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state); +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state);void parse_events_update_lists (struct list_head *list_event,struct list_head *list_all); void parse_events_evlist_error(struct parse_events_state *parse_state,tools/perf/util/parse-events.y +1
-1@@ -199,20 +199,24 @@ group_def group_def: PE_NAME '{' events '}' { - struct list_head *list = $3; + struct list_head *new_head, *list = $3;Quoted 7 lines
tools/perf/util/parse-events.y group_def · line 207
Quoted 5 lines
tools/perf/util/parse-events.y +3
-1inc_group_count(list, _parse_state); - parse_events__set_leader($1, list, _parse_state); + new_head = parse_events__set_leader($1, list, _parse_state); + if (new_head) + list_move(new_head, list);why not put these last 2 lines to
parse_events__set_leaderas well?I can move that, but I'll try to motivate having it this way. The list logic in Linux I find troublesome. In parse-events.y we have a list head that is just a next, prev pointer and not full evsel in a list. So for {topdown-fe-bound,slots} in parse-events.y we have a list of:
head -next->
evsel("topdown-fe-bound")-next->evsel("slots")But in
__perf_evlist__set_leaderthe list doesn't have the head:
evsel("topdown-fe-bound")-next->evsel("slots")The
list_moveis changing the list to:head -next->
evsel("slots")-next->evsel("topdown-fe-bound")which isn't possible in
__perf_evlist__set_leaderas the head has gone. I felt having thelist_moveinparse_events__set_leadermade it look like the leader is being placed second rather than at the head. So this was an attempt to make the code more intention revealing, but that could also be done by some comments, it's just inparse_events__set_leaderwe have the two styles of lists in play.While I'm complaining, there are many brittle assumptions about evlist that it may be worth adding some more assertions for. There is already an assertion that the leader was opened before its siblings. There are assumptions that the idx given to an evsel are incremental and that for the leader 'idx -
prev->idx' will yield the number of list elements. My concern is for libperf users, it doesn't seem unreasonable that they will want to manipulate the evlist but given the assumptions they will likely find bugs - such as idx being used to index an array and reordering breaking things. The metric code would be a customer of reordering were it to need to move slots, but when I experimented with this it easily broke things - which motivates doing it in parse-events or more likely for the metrics, when generating the parse-events string.Thanks,
IanQuoted 20 lines
Quoted 4 lines
tools/perf/util/parse-events.y +0
-0free($1); $$ = list; } |tools/perf/util/parse-events.y group_def · line 219
Quoted 10 lines
tools/perf/util/parse-events.y +4
-2'{' events '}' { - struct list_head *list = $2; + struct list_head *new_head, *list = $2; inc_group_count(list, _parse_state); - parse_events__set_leader(NULL, list, _parse_state); + new_head = parse_events__set_leader(NULL, list, _parse_state); + if (new_head) + list_move(new_head, list);same
thanks,
jirka
kajoljain
Reviewed-byRe: [PATCH 2/2] perf parse-events: Architecture specific leader override
Quoted 56 lines · Ian Rogers
Quoted 47 lines · Ian Rogers
Quoted 45 lines · Jiri Olsa
Quoted 6 lines · Ian Rogers
Currently topdown events must appear after a slots event:
$ perf stat -e '{slots,topdown-fe-bound}' /bin/truePerformance counter stats for '
/bin/true':Quoted 8 lines
3,183,090 slots 986,133 topdown-fe-boundReversing the events yields:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/true Error: The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (topdown-fe-bound).why is this actually failing?
Quoted 14 lines
For metrics the order of events is determined by iterating over a hashmap, and so slots isn't guaranteed to be first which can yield this error.
Change the
set_leaderin parse-events, called when a group is closed, so that rather than always making the first event the leader, if the slots event exists then it is made the leader. It is then moved to the head of the evlist otherwise it won't be opened in the correct order.The result is:
$ perf stat -e '{topdown-fe-bound,slots}' /bin/trueQuoted 8 lines
Performance counter stats for '
/bin/true':3,274,795 slots 1,001,702 topdown-fe-boundA problem with this approach is the slots event is identified by name, names can be overwritten like 'cpu/slots,name=foo/' and this causes the leader change to fail.
would it be better then to move this shuffle to the metric code directly, and make sure it only spits 'good' order of events?
Quoted 7 lines
Quoted 2 lines
and keep "-e '{topdown-fe-bound,slots}'" to fail if user specifies that on the command line
It's an alternative to do that, but the people I spoke to preferred having parse-events do this. I'm not sure what Andi's opinion is. There is a general frustration about how brittle the slots event is, and so this does make it less brittle.
Quoted 4 lines
A different problem this fixes is if you have a group like '{instructions,slots,topdown-fe-bound}' then slots still has to be the group leader even though it is appearing before the topdown event. So before the patch:
Hi Ian,
Thanks for the patch series. We also have similar scenario in case
of powerpc for L2/L3 bus events, where it needs specific event as a group leader. So the solution in this patch series will be helpful in handling it.
Patch series looks good to me.
Thanks,
Kajol Jain
Quoted 207 lines
$ perf stat -e '{instructions,slots,topdown-fe-bound}' -a -- sleep 2Performance counter stats for 'system wide':
<not counted> instructions <not counted> slots <not supported> topdown-fe-bound2.006410898 seconds time elapsedAfter:
$ perf stat -e '{instructions,slots,topdown-fe-bound}' -a -- sleep 2Performance counter stats for 'system wide':
2572632080 slots 362537420 instructions 742416906 topdown-fe-bound2.005875050 seconds time elapsedThanks,
IanQuoted 181 lines
Quoted 111 lines
Quoted 111 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 17 +++++++++++++++++ tools/perf/util/evlist.h | 1 + tools/perf/util/parse-events.c | 16 +++++++++++----- tools/perf/util/parse-events.h | 4 ++-- tools/perf/util/parse-events.y | 12 ++++++++---- 5 files changed, 39 insertions(+), 11 deletions(-)tools/perf/arch/x86/util/evlist.c +17
-0@@ -17,3 +17,20 @@ int arch_evlist__add_default_attrs(struct evlist *evlist)else return parse_events (evlist,TOPDOWN_L1_EVENTS ,NULL ); } + +struct evsel *arch_evlist__leader (struct list_head *list) +{ +struct evsel *evsel, *first; + + first = list_entry(list->next, struct evsel, core.node); + +if (!pmu_have_event ("cpu" ,"slots" )) +return first; + +__evlist__for_each_entry (list, evsel) { +if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && + evsel->name &&strstr (evsel->name,"slots" )) +return evsel; + } +return first; +}tools/perf/util/evlist.h +1
-0@@ -110,6 +110,7 @@ int __evlist__add_default_attrs(struct evlist *evlist, __evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))int arch_evlist__add_default_attrs (struct evlist *evlist); +struct evsel *arch_evlist__leader (struct list_head *list);int evlist__add_dummy (struct evlist *evlist);tools/perf/util/parse-events.c +11
-5@@ -1821,22 +1821,28 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,return ret; } -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state) +__weak struct evsel *arch_evlist__leader (struct list_head *list) +{ + return list_entry(list->next, struct evsel, core.node); +} + +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state) {struct evsel *leader;if (list_empty (list)) {WARN_ONCE (true,"WARNING: failed to set leader: empty list" ); -return ; +return NULL ; }if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state)) -return ; +return NULL ; - leader = list_entry(list->next, struct evsel, core.node); + leader =arch_evlist__leader (list);__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; +return &leader->core.node; }/* list_event is assumed to point to malloc'ed memory */ tools/perf/util/parse-events.h +2
-2@@ -211,8 +211,8 @@ int parse_events_copy_term_list(struct list_head *old,enum perf_pmu_event_symbol_type perf_pmu__parse_check (const char *name); -void parse_events__set_leader (char *name,struct list_head *list, -struct parse_events_state *parse_state); +struct list_head *parse_events__set_leader (char *name,struct list_head *list, +struct parse_events_state *parse_state);void parse_events_update_lists (struct list_head *list_event,struct list_head *list_all); void parse_events_evlist_error(struct parse_events_state *parse_state,tools/perf/util/parse-events.y +1
-1@@ -199,20 +199,24 @@ group_def group_def: PE_NAME '{' events '}' { - struct list_head *list = $3; + struct list_head *new_head, *list = $3;Quoted 7 lines
tools/perf/util/parse-events.y group_def · line 207
Quoted 5 lines
tools/perf/util/parse-events.y +3
-1inc_group_count(list, _parse_state); - parse_events__set_leader($1, list, _parse_state); + new_head = parse_events__set_leader($1, list, _parse_state); + if (new_head) + list_move(new_head, list);why not put these last 2 lines to
parse_events__set_leaderas well?I can move that, but I'll try to motivate having it this way. The list logic in Linux I find troublesome. In parse-events.y we have a list head that is just a next, prev pointer and not full evsel in a list. So for {topdown-fe-bound,slots} in parse-events.y we have a list of:
head -next->
evsel("topdown-fe-bound")-next->evsel("slots")But in
__perf_evlist__set_leaderthe list doesn't have the head:
evsel("topdown-fe-bound")-next->evsel("slots")The
list_moveis changing the list to:head -next->
evsel("slots")-next->evsel("topdown-fe-bound")which isn't possible in
__perf_evlist__set_leaderas the head has gone. I felt having thelist_moveinparse_events__set_leadermade it look like the leader is being placed second rather than at the head. So this was an attempt to make the code more intention revealing, but that could also be done by some comments, it's just inparse_events__set_leaderwe have the two styles of lists in play.While I'm complaining, there are many brittle assumptions about evlist that it may be worth adding some more assertions for. There is already an assertion that the leader was opened before its siblings. There are assumptions that the idx given to an evsel are incremental and that for the leader 'idx -
prev->idx' will yield the number of list elements. My concern is for libperf users, it doesn't seem unreasonable that they will want to manipulate the evlist but given the assumptions they will likely find bugs - such as idx being used to index an array and reordering breaking things. The metric code would be a customer of reordering were it to need to move slots, but when I experimented with this it easily broke things - which motivates doing it in parse-events or more likely for the metrics, when generating the parse-events string.Thanks,
IanQuoted 20 lines
Quoted 4 lines
tools/perf/util/parse-events.y +0
-0free($1); $$ = list; } |tools/perf/util/parse-events.y group_def · line 219
Quoted 10 lines
tools/perf/util/parse-events.y +4
-2'{' events '}' { - struct list_head *list = $2; + struct list_head *new_head, *list = $2; inc_group_count(list, _parse_state); - parse_events__set_leader(NULL, list, _parse_state); + new_head = parse_events__set_leader(NULL, list, _parse_state); + if (new_head) + list_move(new_head, list);same
thanks,
jirka
Patch looks good to me.
Thanks,
Kajol Jain
Quoted 73 lines · Ian Rogers
The leader of a group is the first, but allow it to be an arbitrary list member so that for Intel topdown events slots may always be the group leader.
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/evlist.c | 15 +++++++++------ tools/lib/perf/include/internal/evlist.h | 2 +- tools/perf/util/parse-events.c | 2 +- 3 files changed, 11 insertions(+), 8 deletions(-)tools/lib/perf/evlist.c +9
-6@@ -643,14 +643,14 @@ perf_evlist__next_mmap(struct perf_evlist *evlist, struct perf_mmap *map,return overwrite ? evlist->mmap_ovw_first : evlist->mmap_first; } -void __perf_evlist__set_leader (struct list_head *list) +void __perf_evlist__set_leader (struct list_head *list,struct perf_evsel *leader) { -struct perf_evsel *evsel, *leader; +struct perf_evsel *first, *last, *evsel; - leader = list_entry(list->next, struct perf_evsel, node); - evsel = list_entry(list->prev, struct perf_evsel, node); + first = list_entry(list->next, struct perf_evsel, node); + last = list_entry(list->prev, struct perf_evsel, node); - leader->nr_members = evsel->idx - leader->idx +1 ; + leader->nr_members = last->idx - first->idx +1 ;__perf_evlist__for_each_entry (list, evsel) evsel->leader = leader;@@ -659,7 +659,10 @@ void __perf_evlist__set_leader(struct list_head *list)void perf_evlist__set_leader (struct perf_evlist *evlist) {if (evlist->nr_entries) { +struct perf_evsel *first =list_entry (evlist->entries.next, + struct perf_evsel, node); + evlist->nr_groups = evlist->nr_entries >1 ?1 :0 ; -__perf_evlist__set_leader (&evlist->entries); +__perf_evlist__set_leader (&evlist->entries, first); } }tools/lib/perf/include/internal/evlist.h +1
-1@@ -127,5 +127,5 @@ int perf_evlist__id_add_fd(struct perf_evlist *evlist,void perf_evlist__reset_id_hash (struct perf_evlist *evlist); -void __perf_evlist__set_leader (struct list_head *list); +void __perf_evlist__set_leader (struct list_head *list,struct perf_evsel *leader);#endif /* __LIBPERF_INTERNAL_EVLIST_H */ tools/perf/util/parse-events.c +1
-1@@ -1834,8 +1834,8 @@ void parse_events__set_leader(char *name, struct list_head *list,if (parse_events__set_leader_for_uncore_aliase (name, list, parse_state))return ; -__perf_evlist__set_leader (list); leader = list_entry(list->next, struct evsel, core.node); +__perf_evlist__set_leader (list, &leader->core); leader->group_name = name ?strdup (name) :NULL ; }