Re: [PATCH v1] perf parse-events: Wildcard most "numeric" events
From: Ian Rogers
Date: Wed May 31 2023 - 20:54:56 EST
On Wed, May 31, 2023 at 3:02 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>
> On Wed, May 31, 2023 at 7:04 AM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
> >
> >
> >
> > On 2023-05-30 7:27 p.m., Ian Rogers wrote:
> > > Numeric events are either raw events or those with ABI defined numbers
> > > matched by the lexer. All raw events that don't specify a PMU type
> > > should wildcard match on hybrid systems. So "cycles" should match each
> > > PMU type with an extended type, not just PERF_TYPE_HARDWARE.
> > >
> > > Change wildcard matching to add the event if wildcard PMU scanning
> > > fails, there will be no extended type but this best matches previous
> > > behavior.
> > >
> > > Only set the extended type when the event type supports it and when
> > > perf_pmus__supports_extended_type is true. This new function returns
> > > true if >1 core PMU and avoids potential errors on older kernels.
> > >
> > > Try to always pass a PMU for parse_events_add_numeric, update
> > > evsel__compute_group_pmu_name as software events will have a PMU and
> > > pmu_name rather than NULL. This makes homogeneous and heterogeneous
> > > PMU events more similar.
> > >
> > > Set a parse events error if a hardware term's PMU lookup fails to
> > > provide extra diagnostics.
> >
> > I think the patch tries to fix two issues on hybrid.
> > - Perf fails to create hardware events for each core PMU when the PMU
> > type is not specified.
> > - When splitting a group with multiple PMUs, SW events should be created
> > for each new sub-groups.
>
> Sorry for that. The 2nd item isn't intended. The behavior of software
> events has changed as the code is now assigning all events it can a
> PMU, so software events get the software PMU and a pmu_name of
> "software".
>
> > It works well for the first issue:
> > Before
> >
> > # perf_old stat -e cycles ./hybrid.sh
> >
> > Performance counter stats for './hybrid.sh':
> >
> > 420,833,619 cycles
> >
> > 1.219129587 seconds time elapsed
> >
> > With the patch
> >
> > # perf stat -e cycles ./hybrid.sh
> >
> > Performance counter stats for './hybrid.sh':
> >
> > 420,691,729 cpu_core/cycles/
> > 613,038,584 cpu_atom/cycles/
> > (98.47%)
> >
> > 1.219057759 seconds time elapsed
>
> Thanks for checking.
>
> > For the second issue, the new behavior doesn't seem correct for me.
> >
> > Before
> >
> > ./perf_old stat -e '{data_read,faults}' --no-merge -a sleep 1
> > WARNING: events were regrouped to match PMUs
> > WARNING: grouped events cpus do not match.
> > Events with CPUs not matching the leader will be removed from the group.
> > anon group { data_read, faults }
> >
> > Performance counter stats for 'system wide':
> >
> > 70.05 MiB data_read [uncore_imc_free_running_0]
> > 55.82 MiB data_read [uncore_imc_free_running_1]
> > 77 faults
> >
> > 1.004216885 seconds time elapsed
> >
> >
> > With the patch
> >
> > # ./perf stat -e '{data_read,faults}' --no-merge -a sleep 1
> > WARNING: events were regrouped to match PMUs
> > WARNING: grouped events cpus do not match.
> > Events with CPUs not matching the leader will be removed from the group.
> > anon group { data_read, faults, faults }
> >
> > Performance counter stats for 'system wide':
> >
> > 53.51 MiB data_read [uncore_imc_free_running_0]
> > 39.04 MiB data_read [uncore_imc_free_running_1]
> > 76 cpu_core/faults/
> > 2 cpu_atom/faults/
>
> This is wrong, the patch needs to exclude software events from
> wildcard PMU lookup. It probably makes sense to exclude anything that
> isn't PERF_TYPE_HARDWARE or PERF_TYPE_HW_CACHE.
>
> For PERF_TYPE_RAW, on heterogeneous Intel this matches 'cpu_core' but
> at least for ARM the expectation appears to be that this is all core
> PMUs. Given the confusion, and not extended type, it probably just
> makes sense not to wildcard it.
>
> > The faults is a SW event, but with a core PMU prefix. I think it's wrong.
> >
> > According to the -vvv information, the faults doesn't seem be regrouped
> > with any uncore events.
> >
> > perf_event_attr:
> > type 1
> > size 136
> > config 0x2
> > sample_type IDENTIFIER
> > read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
> > disabled 1
> > inherit 1
> > exclude_guest 1
> > ------------------------------------------------------------
> > sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0 = 5
> > ------------------------------------------------------------
> >
> > For the second issue, can we make the behavior unchanged for now?
>
> Yep. I wasn't aiming to change it.
>
> > BTW: It seems the current perf-tools-next branch has another regression
> > for the JSON event. The inst_retired.any should be available for both
> > atom and core. I believe I tried the below commands several days ago.
> > But it does't work now. I will check when the regression was introduced.
> >
> > # ./perf stat -e inst_retired.any ./hybrid.sh
> >
> > Performance counter stats for './hybrid.sh':
> >
> > 3,261,510,386 cpu_core/inst_retired.any/
> > (99.07%)
> >
> > 1.218858763 seconds time elapsed
>
> Thanks, this was supposed to be unchanged for this patch. The matching
> logic is here:
> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/parse-events.y?h=perf-tools-next#n284
>
> I have access to an Alderlake again now, so I'll test the patch and do
> the type changes mentioned above for v2.
This was caused by commit 6b9da2607030 ("perf pmu: Remove
is_pmu_hybrid") specifically:
```
bool perf_pmu__auto_merge_stats(const struct perf_pmu *pmu)
{
- return !is_pmu_hybrid(pmu->name);
+ return pmu->is_core && perf_pmus__num_core_pmus() > 1;
```
where the sense of the test was flipped causing the counter values to
be merged. I'll send the fix.
Thanks,
Ian
> Thanks,
> Ian
>
> > Thanks,
> > Kan
> >
> > >
> > > Reported-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > > Fixes: 8bc75f699c14 ("perf parse-events: Support wildcards on raw events")
> > > Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>> ---
> > > tools/perf/util/parse-events.c | 71 +++++++++++++++++++++-------------
> > > tools/perf/util/parse-events.y | 4 +-
> > > tools/perf/util/pmus.c | 5 +++
> > > tools/perf/util/pmus.h | 1 +
> > > 4 files changed, 53 insertions(+), 28 deletions(-)
> > >
> > > diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> > > index 7f047ac11168..58c9f34bcd95 100644
> > > --- a/tools/perf/util/parse-events.c
> > > +++ b/tools/perf/util/parse-events.c
> > > @@ -372,7 +372,7 @@ static int config_attr(struct perf_event_attr *attr,
> > > * contain hyphens and the longest name
> > > * should always be selected.
> > > */
> > > -int parse_events__decode_legacy_cache(const char *name, int pmu_type, __u64 *config)
> > > +int parse_events__decode_legacy_cache(const char *name, int extended_pmu_type, __u64 *config)
> > > {
> > > int len, cache_type = -1, cache_op = -1, cache_result = -1;
> > > const char *name_end = &name[strlen(name) + 1];
> > > @@ -423,8 +423,9 @@ int parse_events__decode_legacy_cache(const char *name, int pmu_type, __u64 *con
> > > if (cache_result == -1)
> > > cache_result = PERF_COUNT_HW_CACHE_RESULT_ACCESS;
> > >
> > > - *config = ((__u64)pmu_type << PERF_PMU_TYPE_SHIFT) |
> > > - cache_type | (cache_op << 8) | (cache_result << 16);
> > > + *config = cache_type | (cache_op << 8) | (cache_result << 16);
> > > + if (perf_pmus__supports_extended_type())
> > > + *config |= (__u64)extended_pmu_type << PERF_PMU_TYPE_SHIFT;
> > > return 0;
> > > }
> > >
> > > @@ -1204,11 +1205,17 @@ static int config_term_pmu(struct perf_event_attr *attr,
> > > const struct perf_pmu *pmu = perf_pmus__find_by_type(attr->type);
> > >
> > > if (!pmu) {
> > > - pr_debug("Failed to find PMU for type %d", attr->type);
> > > + char *err_str;
> > > +
> > > + if (asprintf(&err_str, "Failed to find PMU for type %d", attr->type) >= 0)
> > > + parse_events_error__handle(err, term->err_term,
> > > + err_str, /*help=*/NULL);
> > > return -EINVAL;
> > > }
> > > attr->type = PERF_TYPE_HARDWARE;
> > > - attr->config = ((__u64)pmu->type << PERF_PMU_TYPE_SHIFT) | term->val.num;
> > > + attr->config = term->val.num;
> > > + if (perf_pmus__supports_extended_type())
> > > + attr->config |= (__u64)pmu->type << PERF_PMU_TYPE_SHIFT;
> > > return 0;
> > > }
> > > if (term->type_term == PARSE_EVENTS__TERM_TYPE_USER ||
> > > @@ -1435,8 +1442,8 @@ int parse_events_add_tracepoint(struct list_head *list, int *idx,
> > >
> > > static int __parse_events_add_numeric(struct parse_events_state *parse_state,
> > > struct list_head *list,
> > > - struct perf_pmu *pmu, u32 type, u64 config,
> > > - struct list_head *head_config)
> > > + struct perf_pmu *pmu, u32 type, u32 extended_type,
> > > + u64 config, struct list_head *head_config)
> > > {
> > > struct perf_event_attr attr;
> > > LIST_HEAD(config_terms);
> > > @@ -1446,6 +1453,10 @@ static int __parse_events_add_numeric(struct parse_events_state *parse_state,
> > > memset(&attr, 0, sizeof(attr));
> > > attr.type = type;
> > > attr.config = config;
> > > + if (extended_type && (type == PERF_TYPE_HARDWARE || type == PERF_TYPE_HW_CACHE)) {
> > > + assert(perf_pmus__supports_extended_type());
> > > + attr.config |= (u64)extended_type << PERF_PMU_TYPE_SHIFT;
> > > + };
> > >
> > > if (head_config) {
> > > if (config_attr(&attr, head_config, parse_state->error,
> > > @@ -1474,24 +1485,26 @@ int parse_events_add_numeric(struct parse_events_state *parse_state,
> > > struct perf_pmu *pmu = NULL;
> > > bool found_supported = false;
> > >
> > > - if (!wildcard)
> > > - return __parse_events_add_numeric(parse_state, list, /*pmu=*/NULL,
> > > - type, config, head_config);
> > > -
> > > /* Wildcards on numeric values are only supported by core PMUs. */
> > > - while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
> > > - int ret;
> > > + if (wildcard && perf_pmus__supports_extended_type()) {
> > > + while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
> > > + int ret;
> > >
> > > - if (parse_events__filter_pmu(parse_state, pmu))
> > > - continue;
> > > + found_supported = true;
> > > + if (parse_events__filter_pmu(parse_state, pmu))
> > > + continue;
> > >
> > > - found_supported = true;
> > > - ret = __parse_events_add_numeric(parse_state, list, pmu, pmu->type,
> > > - config, head_config);
> > > - if (ret)
> > > - return ret;
> > > + ret = __parse_events_add_numeric(parse_state, list, pmu,
> > > + type, pmu->type,
> > > + config, head_config);
> > > + if (ret)
> > > + return ret;
> > > + }
> > > + if (found_supported)
> > > + return 0;
> > > }
> > > - return found_supported ? 0 : -EINVAL;
> > > + return __parse_events_add_numeric(parse_state, list, perf_pmus__find_by_type(type),
> > > + type, /*extended_type=*/0, config, head_config);
> > > }
> > >
> > > int parse_events_add_tool(struct parse_events_state *parse_state,
> > > @@ -1989,7 +2002,10 @@ static int evsel__compute_group_pmu_name(struct evsel *evsel,
> > > {
> > > struct evsel *leader = evsel__leader(evsel);
> > > struct evsel *pos;
> > > - const char *group_pmu_name = evsel->pmu_name ?: "cpu";
> > > + const char *group_pmu_name = "cpu";
> > > +
> > > + if (evsel->core.attr.type != PERF_TYPE_SOFTWARE && evsel->pmu_name)
> > > + group_pmu_name = evsel->pmu_name;
> > >
> > > /*
> > > * Software events may be in a group with other uncore PMU events. Use
> > > @@ -2002,14 +2018,17 @@ static int evsel__compute_group_pmu_name(struct evsel *evsel,
> > > if (evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) {
> > > /*
> > > * Starting with the leader, find the first event with a named
> > > - * PMU. for_each_group_(member|evsel) isn't used as the list
> > > - * isn't yet sorted putting evsel's in the same group together.
> > > + * non-software PMU. for_each_group_(member|evsel) isn't used as
> > > + * the list isn't yet sorted putting evsel's in the same group
> > > + * together.
> > > */
> > > - if (leader->pmu_name) {
> > > + if (leader->pmu_name && leader->core.attr.type != PERF_TYPE_SOFTWARE) {
> > > group_pmu_name = leader->pmu_name;
> > > } else if (leader->core.nr_members > 1) {
> > > list_for_each_entry(pos, head, core.node) {
> > > - if (evsel__leader(pos) == leader && pos->pmu_name) {
> > > + if (evsel__leader(pos) == leader &&
> > > + pos->core.attr.type != PERF_TYPE_SOFTWARE &&
> > > + pos->pmu_name) {
> > > group_pmu_name = pos->pmu_name;
> > > break;
> > > }
> > > diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
> > > index abd6ab460e12..48c21314cf6b 100644
> > > --- a/tools/perf/util/parse-events.y
> > > +++ b/tools/perf/util/parse-events.y
> > > @@ -449,7 +449,7 @@ value_sym '/' event_config '/'
> > > list = alloc_list();
> > > ABORT_ON(!list);
> > > err = parse_events_add_numeric(_parse_state, list, type, config, $3,
> > > - /*wildcard=*/false);
> > > + /*wildcard=*/true);
> > > parse_events_terms__delete($3);
> > > if (err) {
> > > free_list_evsel(list);
> > > @@ -468,7 +468,7 @@ value_sym sep_slash_slash_dc
> > > ABORT_ON(!list);
> > > ABORT_ON(parse_events_add_numeric(_parse_state, list, type, config,
> > > /*head_config=*/NULL,
> > > - /*wildcard=*/false));
> > > + /*wildcard=*/true));
> > > $$ = list;
> > > }
> > > |
> > > diff --git a/tools/perf/util/pmus.c b/tools/perf/util/pmus.c
> > > index 53f11f6ce878..e1d0a93147e5 100644
> > > --- a/tools/perf/util/pmus.c
> > > +++ b/tools/perf/util/pmus.c
> > > @@ -477,6 +477,11 @@ int perf_pmus__num_core_pmus(void)
> > > return count;
> > > }
> > >
> > > +bool perf_pmus__supports_extended_type(void)
> > > +{
> > > + return perf_pmus__num_core_pmus() > 1;
> > > +}
> > > +
> > > struct perf_pmu *evsel__find_pmu(const struct evsel *evsel)
> > > {
> > > struct perf_pmu *pmu = evsel->pmu;
> > > diff --git a/tools/perf/util/pmus.h b/tools/perf/util/pmus.h
> > > index 1e710720aec7..d02ffea5d3a4 100644
> > > --- a/tools/perf/util/pmus.h
> > > +++ b/tools/perf/util/pmus.h
> > > @@ -19,5 +19,6 @@ int perf_pmus__num_mem_pmus(void);
> > > void perf_pmus__print_pmu_events(const struct print_callbacks *print_cb, void *print_state);
> > > bool perf_pmus__have_event(const char *pname, const char *name);
> > > int perf_pmus__num_core_pmus(void);
> > > +bool perf_pmus__supports_extended_type(void);
> > >
> > > #endif /* __PMUS_H */