Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Yang Jihong <yangjihong1@huawei.com>
To:     <peterz@infradead.org>, <mingo@redhat.com>, <acme@kernel.org>,
        <mark.rutland@arm.com>, <alexander.shishkin@linux.intel.com>,
        <jolsa@kernel.org>, <namhyung@kernel.org>, <irogers@google.com>,
        <adrian.hunter@intel.com>, <kan.liang@linux.intel.com>,
        <james.clark@arm.com>, <tmricht@linux.ibm.com>,
        <ak@linux.intel.com>, <anshuman.khandual@arm.com>,
        <linux-kernel@vger.kernel.org>, <linux-perf-users@vger.kernel.org>
CC:     <yangjihong1@huawei.com>
Subject: [PATCH v4 4/7] perf record: Track sideband events for all CPUs when tracing selected CPUs
Date:   Wed, 2 Aug 2023 07:49:45 +0000
Message-ID: <20230802074948.136468-5-yangjihong1@huawei.com>
In-Reply-To: <20230802074948.136468-1-yangjihong1@huawei.com>
References: <20230802074948.136468-1-yangjihong1@huawei.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 7BIT
Content-Type:   text/plain; charset=US-ASCII
Precedence: bulk

User space tasks can migrate between CPUs, we need to track side-band
events for all CPUs.

The specific scenarios are as follows:

         CPU0                                 CPU1
  perf record -C 0 start
                              taskA starts to be created and executed
                                -> PERF_RECORD_COMM and PERF_RECORD_MMAP
                                   events only deliver to CPU1
                              ......
                                |
                          migrate to CPU0
                                |
  Running on CPU0    <----------/
  ...

  perf record -C 0 stop

Now perf samples the PC of taskA. However, perf does not record the
PERF_RECORD_COMM and PERF_RECORD_MMAP events of taskA.
Therefore, the comm and symbols of taskA cannot be parsed.

The solution is to record sideband events for all CPUs when tracing
selected CPUs. Because this modifies the default behavior, add related
comments to the perf record man page.

The sys_perf_event_open invoked is as follows:

  # perf --debug verbose=3 record -e cpu-clock -C 1 true
  <SNIP>
  Opening: cpu-clock
  ------------------------------------------------------------
  perf_event_attr:
    type                             1 (PERF_TYPE_SOFTWARE)
    size                             136
    config                           0 (PERF_COUNT_SW_CPU_CLOCK)
    { sample_period, sample_freq }   4000
    sample_type                      IP|TID|TIME|CPU|PERIOD|IDENTIFIER
    read_format                      ID|LOST
    disabled                         1
    inherit                          1
    freq                             1
    sample_id_all                    1
    exclude_guest                    1
  ------------------------------------------------------------
  sys_perf_event_open: pid -1  cpu 1  group_fd -1  flags 0x8 = 5
  Opening: dummy:u
  ------------------------------------------------------------
  perf_event_attr:
    type                             1 (PERF_TYPE_SOFTWARE)
    size                             136
    config                           0x9 (PERF_COUNT_SW_DUMMY)
    { sample_period, sample_freq }   1
    sample_type                      IP|TID|TIME|CPU|IDENTIFIER
    read_format                      ID|LOST
    inherit                          1
    exclude_kernel                   1
    exclude_hv                       1
    mmap                             1
    comm                             1
    task                             1
    sample_id_all                    1
    exclude_guest                    1
    mmap2                            1
    comm_exec                        1
    ksymbol                          1
    bpf_event                        1
  ------------------------------------------------------------
  sys_perf_event_open: pid -1  cpu 0  group_fd -1  flags 0x8 = 6
  sys_perf_event_open: pid -1  cpu 1  group_fd -1  flags 0x8 = 7
  sys_perf_event_open: pid -1  cpu 2  group_fd -1  flags 0x8 = 9
  sys_perf_event_open: pid -1  cpu 3  group_fd -1  flags 0x8 = 10
  sys_perf_event_open: pid -1  cpu 4  group_fd -1  flags 0x8 = 11
  sys_perf_event_open: pid -1  cpu 5  group_fd -1  flags 0x8 = 12
  sys_perf_event_open: pid -1  cpu 6  group_fd -1  flags 0x8 = 13
  sys_perf_event_open: pid -1  cpu 7  group_fd -1  flags 0x8 = 14
  <SNIP>

Signed-off-by: Yang Jihong <yangjihong1@huawei.com>
---
 tools/perf/Documentation/perf-record.txt |  3 ++
 tools/perf/builtin-record.c              | 45 +++++++++++++++++++++++-
 2 files changed, 47 insertions(+), 1 deletion(-)

diff --git a/tools/perf/Documentation/perf-record.txt b/tools/perf/Documentation/perf-record.txt
index 680396c56bd1..dac53ece51ab 100644
--- a/tools/perf/Documentation/perf-record.txt
+++ b/tools/perf/Documentation/perf-record.txt
@@ -388,6 +388,9 @@ comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-
 In per-thread mode with inheritance mode on (default), samples are captured only when
 the thread executes on the designated CPUs. Default is to monitor all CPUs.
 
+User space tasks can migrate between CPUs, so when tracing selected CPUs,
+a dummy event is created to track sideband for all CPUs.
+
 -B::
 --no-buildid::
 Do not save the build ids of binaries in the perf.data files. This skips
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 3ff9d972225e..468afeaac52d 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -908,10 +908,45 @@ static int record__config_off_cpu(struct record *rec)
 	return off_cpu_prepare(rec->evlist, &rec->opts.target, &rec->opts);
 }
 
+static bool record__tracking_system_wide(struct record *rec)
+{
+	struct record_opts *opts = &rec->opts;
+	struct evlist *evlist = rec->evlist;
+	struct evsel *evsel;
+
+	/*
+	 * If all (non-dummy) evsel have exclude_user,
+	 * system_wide is not needed.
+	 *
+	 * all_kernel and all_user will overwrite exclude_kernel and
+	 * exclude_user of attr in evsel__config(), here need to check
+	 * all the three items.
+	 *
+	 * Sideband system wide if one of the following conditions is met:
+	 *
+	 *   - all_user is set, and there is a non-dummy event
+	 *   - all_user and all_kernel are not set, and there is
+	 *     a non-dummy event without exclude_user
+	 */
+	evlist__for_each_entry(evlist, evsel) {
+		if (!evsel__is_dummy_event(evsel)) {
+			if (opts->all_user)
+				return true;
+
+			if (!opts->all_user && !opts->all_kernel &&
+			    !evsel->core.attr.exclude_user)
+				return true;
+		}
+	}
+
+	return false;
+}
+
 static int record__config_tracking_events(struct record *rec)
 {
 	struct record_opts *opts = &rec->opts;
 	struct evlist *evlist = rec->evlist;
+	bool system_wide = false;
 	struct evsel *evsel;
 
 	/*
@@ -921,7 +956,15 @@ static int record__config_tracking_events(struct record *rec)
 	 */
 	if (opts->target.initial_delay || target__has_cpu(&opts->target) ||
 	    perf_pmus__num_core_pmus() > 1) {
-		evsel = evlist__findnew_tracking_event(evlist, false);
+
+		/*
+		 * User space tasks can migrate between CPUs, so when tracing
+		 * selected CPUs, sideband for all CPUs is still needed.
+		 */
+		if (!!opts->target.cpu_list && record__tracking_system_wide(rec))
+			system_wide = true;
+
+		evsel = evlist__findnew_tracking_event(evlist, system_wide);
 		if (!evsel)
 			return -ENOMEM;
 
-- 
2.30.GIT