Thread (5 messages) flat view 5 messages, 2 authors, 2021-08-25

Re: [Question] perf tools: lex parsing issue

From: Ian Rogers <irogers@google.com>
Date: 2021-08-25 00:14:08
Also in: lkml
Subsystem: performance events subsystem, the rest · Maintainers: Peter Zijlstra, Ingo Molnar, Arnaldo Carvalho de Melo, Namhyung Kim, Linus Torvalds

On Mon, Aug 23, 2021 at 4:23 AM John Garry [off-list ref] wrote:
quoted hunk ↗ jump to hunk
Hi jirka,

If you remember from some time ago we discussed how the lex parsing
creates strange aliases:

https://lore.kernel.org/lkml/20200320093006.GA1343171@krava/ (local)

I am no expert on l+y, but it seems that we simply don't set the term
config field for known term types. Well, not for
PARSE_EVENTS__TERM_TYPE_SAMPLE_PERIOD type anyway.

This super hack resolves that issue:

--->8----
--- a/tools/perf/util/parse-events.y
+++ b/tools/perf/util/parse-events.y
@@ -765,7 +765,12 @@ event_config ',' event_term
struct list_head *head = $1;
struct parse_events_term *term = $3

+ if (term->type_term == PARSE_EVENTS__TERM_TYPE_SAMPLE_PERIOD) {
+       term->config = strdup("period");
+ }
+
if (!head) {
        parse_events_term__delete(term);
        YYABORT;
--

----8-----
Agreed this is hacky, I think it'd be better to fix this up in the
output. For example:
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
index 6cdbee8a12e7..c77c42275efa 100644
--- a/tools/perf/util/pmu.c
+++ b/tools/perf/util/pmu.c
@@ -365,15 +365,21 @@ static int __perf_pmu__new_alias(struct
list_head *list, char *dir, char *name,
        memset(newval, 0, sizeof(newval));
        ret = 0;
        list_for_each_entry(term, &alias->terms, list) {
+               const char * config = term->config;
+
                if (ret)
                        ret += scnprintf(newval + ret, sizeof(newval) - ret,
                                         ",");
+               if (!config) {
+                       /* Note: config_term_names in parse_events.c
isn't accessible */
+                       config = config_term_names[term->type_term];
+               }
                if (term->type_val == PARSE_EVENTS__TERM_TYPE_NUM)
                        ret += scnprintf(newval + ret, sizeof(newval) - ret,
-                                        "%s=%#x", term->config, term->val.num);
+                                        "%s=%#x", config, term->val.num);
                else if (term->type_val == PARSE_EVENTS__TERM_TYPE_STR)
                        ret += scnprintf(newval + ret, sizeof(newval) - ret,
-                                        "%s=%s", term->config, term->val.str);
+                                        "%s=%s", config, term->val.str);
        }

        alias->name = strdup(name);

It looks like a similar fix is needed in format_alias.

Thanks,
Ian
So we get "umask=0x80,period=0x30d40,event=0x6" now, rather than
"umask=0x80,(null)=0x30d40,event=0x6", for the perf_pmu_alias.str, as an
example.

Did you ever get a chance to look into this issue? Do you know how could
or should this field be set properly?

Some more background:
The reason I was looking at this is because I think it causes a problem
for pmu-events (JSONs) aliasing for some PMUs. Specifically it's PMU
which use "config=xxx" in sysfs files in
/sys/bus/event_source/devices/PMUx/events/, rather than "event=xxx". The
actual problem is that I trigger this warn in pmu.c:

static void perf_pmu_assign_str(char *name, const char *field, char
**old_str,
char **new_str)
{

if (*new_str) { /* Have new string, check with old */
        if (strcasecmp(*old_str, *new_str))
                pr_debug("alias %s differs i ... <---

As I get "config=event=0xXXX" vs "config=(null)=0xXXX"

As I am not sure how to solve that yet, but, since we have
config=(null), I thought it best to solve the first issue first.

Thanks,
John
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help