Re: [PATCH 1/2] perf lock report: Restore aggregation by caller by default

From: Nick Forrington
Date: Fri Nov 10 2023 - 12:42:21 EST



On 02/11/2023 05:55, Namhyung Kim wrote:
Hello,

On Tue, Oct 31, 2023 at 5:05 AM Nick Forrington <nick.forrington@xxxxxxx> wrote:
This change restores the previous default behaviour for "perf lock
report", making the current aggregate-by-address behaviour available via
the new "--lock-addr" command line parameter.

This makes the behaviour consistent with "perf lock contention" (which
also aggregates by caller by default, or by address when "--lock-addr"
is specified).
I understand your concern but actually there's a difference.
"perf lock contention" is a new command which works with new
contention tracepoints whereas "perf lock report" works with old
lockdep/lockstat tracepoints which are not available in the default
configuration.

I made "perf lock contention" compatible to "perf lock report" so
it mimics the old tracepoints behavior as much as possible using
new tracepoints. But the important difference is that new contention
tracepoints don't have lock names. The old perf lock report showed
lock names by default but contention output had to use the caller
instead.

Thanks for the information.


It looks like "perf lock record" requests both tracepoint types.


However, on the system I used the lockdep tracepoints were not available - only lock:contention_begin/end.


Commit 688d2e8de231 ("perf lock contention: Add -l/--lock-addr option")
introduced aggregation modes for "perf lock contention" and (potentially
inadvertently) changed the behaviour of "perf lock report" from
aggregate-by-caller to aggregate-by-address (making the prior behaviour
inaccessible).
So it doesn't change the behavior of perf lock report.
You're adding a new (default) feature for perf lock report
to sort the output by caller. And please note that caller
info needs callstacks. perf lock record adds it by default
when it finds there are only lock contention tracepoints.
But if it really has the old tracepoints, caller won't work
unless you enabled callstack collection manually (-g).

With the caveat that only lock:contention_begin/end were available in my test, there does appear to be a difference in output for perf lock report.


I arrived at 688d2e8de231 via git bisect. 688d2e8de231~ (eca949b2b4ad) shows the aggregate-by-caller behaviour (seen below, and "restored" by this patch).


Example aggregate-by-address output:

$ perf lock report -F acquired
I guess you need -l option here.

This was the default output before this patch (or output with the patch and "-l"), I can clarify this if the patch goes ahead.

Name acquired

event_mutex 34
21
1
This is because you used contention tracepoints
and they don't have lock names.

Yes, the lockdep tracepoints weren't available for my test, although on the same machine, versions prior to 688d2e8de231 appear to show the below output rather than the above.


Example aggregate-by-caller output:

$ perf lock report -F acquired
Name acquired

perf_trace_init+... 34
lock_mm_and_find... 20
inherit_event.co... 1
do_madvise+0x1f8 1
Maybe it's ok to change the default behavior for contention
tracepoints. But when lockdep tracepoints are available
it should use the existing addr (symbol) mode.

Cc: stable@xxxxxxxxxx
Fixes: 688d2e8de231 ("perf lock contention: Add -l/--lock-addr option")
So, I don't think this is a fix.

Thanks,
Namhyung


Thanks,
Nick



Signed-off-by: Nick Forrington <nick.forrington@xxxxxxx>
---
tools/perf/Documentation/perf-lock.txt | 4 ++++
tools/perf/builtin-lock.c | 24 +++++++++++++++++++++---
2 files changed, 25 insertions(+), 3 deletions(-)

diff --git a/tools/perf/Documentation/perf-lock.txt b/tools/perf/Documentation/perf-lock.txt
index 503abcba1438..349333acbbfc 100644
--- a/tools/perf/Documentation/perf-lock.txt
+++ b/tools/perf/Documentation/perf-lock.txt
@@ -80,6 +80,10 @@ REPORT OPTIONS
--combine-locks::
Merge lock instances in the same class (based on name).

+-l::
+--lock-addr::
+ Show lock contention stat by address
+
-t::
--threads::
The -t option is to show per-thread lock stat like below:
diff --git a/tools/perf/builtin-lock.c b/tools/perf/builtin-lock.c
index fa7419978353..3aa8ba5ad928 100644
--- a/tools/perf/builtin-lock.c
+++ b/tools/perf/builtin-lock.c
@@ -78,7 +78,7 @@ struct callstack_filter {

static struct lock_filter filters;

-static enum lock_aggr_mode aggr_mode = LOCK_AGGR_ADDR;
+static enum lock_aggr_mode aggr_mode = LOCK_AGGR_CALLER;

static bool needs_callstack(void)
{
@@ -1983,8 +1983,8 @@ static int __cmd_report(bool display_info)
if (select_key(false))
goto out_delete;

- if (show_thread_stats)
- aggr_mode = LOCK_AGGR_TASK;
+ aggr_mode = show_thread_stats ? LOCK_AGGR_TASK :
+ show_lock_addrs ? LOCK_AGGR_ADDR : LOCK_AGGR_CALLER;

err = perf_session__process_events(session);
if (err)
@@ -2008,6 +2008,19 @@ static void sighandler(int sig __maybe_unused)
{
}

+static int check_lock_report_options(const struct option *options,
+ const char * const *usage)
+{
+ if (show_thread_stats && show_lock_addrs) {
+ pr_err("Cannot use thread and addr mode together\n");
+ parse_options_usage(usage, options, "threads", 0);
+ parse_options_usage(NULL, options, "lock-addr", 0);
+ return -1;
+ }
+
+ return 0;
+}
+
static int check_lock_contention_options(const struct option *options,
const char * const *usage)

@@ -2589,6 +2602,7 @@ int cmd_lock(int argc, const char **argv)
/* TODO: type */
OPT_BOOLEAN('c', "combine-locks", &combine_locks,
"combine locks in the same class"),
+ OPT_BOOLEAN('l', "lock-addr", &show_lock_addrs, "show lock stats by address"),
OPT_BOOLEAN('t', "threads", &show_thread_stats,
"show per-thread lock stats"),
OPT_INTEGER('E', "entries", &print_nr_entries, "display this many functions"),
@@ -2680,6 +2694,10 @@ int cmd_lock(int argc, const char **argv)
if (argc)
usage_with_options(report_usage, report_options);
}
+
+ if (check_lock_report_options(report_options, report_usage) < 0)
+ return -1;
+
rc = __cmd_report(false);
} else if (!strcmp(argv[0], "script")) {
/* Aliased to 'perf script' */
--
2.42.0