Merge tag 'perf-tools-for-v6.13-2024-11-24' of git://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools

Pull perf tools updates from Namhyung Kim:
 "perf record:

   - Enable leader sampling for inherited task events. It was supported
     only for system-wide events but the kernel started to support such
     a setup since v6.12.

     This is to reduce the number of PMU interrupts. The samples of the
     leader event will contain counts of other events and no samples
     will be generated for the other member events.

       $ perf record -e '{cycles,instructions}:S'  ${MYPROG}

  perf report:

   - Fix --branch-history option to display more branch-related
     information like prediction, abort and cycles which is available
     on Intel machines.

       $ perf record -bg -- perf test -w brstack

       $ perf report --branch-history
       ...
       #
       # Overhead  Source:Line               Symbol          Shared Object         Predicted  Abort  Cycles  IPC   [IPC Coverage]
       # ........  ........................  ..............  ....................  .........  .....  ......  ....................
       #
            8.17%  copy_page_64.S:19         [k] copy_page   [kernel.kallsyms]     50.0%      0      5       -      -
                   |
                   ---xas_load xarray.h:171
                      |
                      |--5.68%--xas_load xarray.c:245 (cycles:1)
                      |          xas_load xarray.c:242
                      |          xas_load xarray.h:1260 (cycles:1)
                      |          xas_descend xarray.c:146
                      |          xas_load xarray.c:244 (cycles:2)
                      |          xas_load xarray.c:245
                      |          xas_descend xarray.c:218 (cycles:10)
       ...

  perf stat:

   - Add HWMON PMU support.

     The HWMON provides various system information like CPU/GPU
     temperature, fan speed and so on. Expose them as PMU events so that
     users can see the values using perf stat commands.

       $ perf stat -e temp_cpu,fan1 true

        Performance counter stats for 'true':

                    60.00 'C   temp_cpu
                        0 rpm  fan1

              0.000745382 seconds time elapsed

              0.000883000 seconds user
              0.000000000 seconds sys

   - Display metric threshold in JSON output.

     Some metrics define thresholds to classify value ranges. It used to
     be in a different color but it won't work for JSON.

     Add "metric-threshold" field to the JSON that can be one of "good",
     "less good", "nearly bad" and "bad".

       # perf stat -a -M TopdownL1 -j true
       {"counter-value" : "18693525.000000", "unit" : "", "event" : "TOPDOWN.SLOTS", "event-runtime" : 5552708, "pcnt-running" : 100.00, "metric-value" : "43.226002", "metric-unit" : "%  tma_backend_bound", "metric-threshold" : "bad"}
       {"metric-value" : "29.212267", "metric-unit" : "%  tma_frontend_bound", "metric-threshold" : "bad"}
       {"metric-value" : "7.138972", "metric-unit" : "%  tma_bad_speculation", "metric-threshold" : "good"}
       {"metric-value" : "20.422759", "metric-unit" : "%  tma_retiring", "metric-threshold" : "good"}
       {"counter-value" : "3817732.000000", "unit" : "", "event" : "topdown-retiring", "event-runtime" : 5552708, "pcnt-running" : 100.00, }
       {"counter-value" : "5472824.000000", "unit" : "", "event" : "topdown-fe-bound", "event-runtime" : 5552708, "pcnt-running" : 100.00, }
       {"counter-value" : "7984780.000000", "unit" : "", "event" : "topdown-be-bound", "event-runtime" : 5552708, "pcnt-running" : 100.00, }
       {"counter-value" : "1418181.000000", "unit" : "", "event" : "topdown-bad-spec", "event-runtime" : 5552708, "pcnt-running" : 100.00, }
       ...

  perf sched:

   - Add -P/--pre-migrations option for 'timehist' sub-command to track
     time a task waited on a run-queue before migrating to a different
     CPU.

       $ perf sched timehist -P
                  time    cpu  task name                       wait time  sch delay   run time  pre-mig time
                               [tid/pid]                          (msec)     (msec)     (msec)     (msec)
       --------------- ------  ------------------------------  ---------  ---------  ---------  ---------
         585940.535527 [0000]  perf[584885]                        0.000      0.000      0.000      0.000
         585940.535535 [0000]  migration/0[20]                     0.000      0.002      0.008      0.000
         585940.535559 [0001]  perf[584885]                        0.000      0.000      0.000      0.000
         585940.535563 [0001]  migration/1[25]                     0.000      0.001      0.004      0.000
         585940.535678 [0002]  perf[584885]                        0.000      0.000      0.000      0.000
         585940.535686 [0002]  migration/2[31]                     0.000      0.002      0.008      0.000
         585940.535905 [0001]  <idle>                              0.000      0.000      0.342      0.000
         585940.535938 [0003]  perf[584885]                        0.000      0.000      0.000      0.000
         585940.537048 [0001]  sleep[584886]                       0.000      0.019      1.142      0.001
         585940.537749 [0002]  <idle>                              0.000      0.000      2.062      0.000
       ...

  Build:

   - Make libunwind opt-in (LIBUNWIND=1) rather than opt-out.

     The perf tools are generally built with libelf and libdw which has
     unwinder functionality. The libunwind support predates it and no
     need to have duplicate unwinders by default.

   - Rename NO_DWARF=1 build option to NO_LIBDW=1 in order to clarify
     it's using libdw for handling DWARF information.

  Internals:

   - Do not set exclude_guest bit in the perf_event_attr by default.

     This was causing a trouble in AMD IBS PMU as it doesn't support the
     bit. The bit will be set when it's needed later by the fallback
     logic. Also update the missing feature detection logic to make sure
     not clear supported bits unnecessarily.

   - Run perf test in parallel by default and mark flaky tests
     "exclusive" to run them serially at the end. Some test numbers are
     changed but the test can complete in less than half the time.

  JSON vendor events:

   - Add AMD Zen 5 events and metrics.

   - Add i.MX91 and i.MX95 DDR metrics

   - Fix HiSilicon HIP08 Topdown metric name.

   - Support compat events on PowerPC"

* tag 'perf-tools-for-v6.13-2024-11-24' of git://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools: (232 commits)
  perf tests: Fix hwmon parsing with PMU name test
  perf hwmon_pmu: Ensure hwmon key union is zeroed before use
  perf tests hwmon_pmu: Remove double evlist__delete()
  perf/test: fix perf ftrace test on s390
  perf bpf-filter: Return -ENOMEM directly when pfi allocation fails
  perf test: Correct hwmon test PMU detection
  perf: Remove unused del_perf_probe_events()
  perf pmu: Move pmu_metrics_table__find and remove ARM override
  perf jevents: Add map_for_cpu()
  perf header: Pass a perf_cpu rather than a PMU to get_cpuid_str
  perf header: Avoid transitive PMU includes
  perf arm64 header: Use cpu argument in get_cpuid
  perf header: Refactor get_cpuid to take a CPU for ARM
  perf header: Move is_cpu_online to numa bench
  perf jevents: fix breakage when do perf stat on system metric
  perf test: Add missing __exit calls in tool/hwmon tests
  perf tests: Make leader sampling test work without branch event
  perf util: Remove kernel version deadcode
  perf test shell trace_exit_race: Use --no-comm to avoid cases where COMM isn't resolved
  perf test shell trace_exit_race: Show what went wrong in verbose mode
  ...

1 files changed, 91 insertions, 7 deletions

diff --git a/tools/perf/util/bpf_skel/bperf_follower.bpf.c b/tools/perf/util/bpf_skel/bperf_follower.bpf.c
index f193998530d4..0595063139a3 100644
--- a/tools/perf/util/bpf_skel/bperf_follower.bpf.c
+++ b/tools/perf/util/bpf_skel/bperf_follower.bpf.c
@@ -5,6 +5,8 @@
 #include <bpf/bpf_tracing.h>
 #include "bperf_u.h"
 
+#define MAX_ENTRIES 102400
+
 struct {
 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
 	__uint(key_size, sizeof(__u32));
@@ -22,25 +24,29 @@ struct {
 struct {
 	__uint(type, BPF_MAP_TYPE_HASH);
 	__uint(key_size, sizeof(__u32));
-	__uint(value_size, sizeof(__u32));
+	__uint(value_size, sizeof(struct bperf_filter_value));
+	__uint(max_entries, MAX_ENTRIES);
+	__uint(map_flags, BPF_F_NO_PREALLOC);
 } filter SEC(".maps");
 
 enum bperf_filter_type type = 0;
 int enabled = 0;
+int inherit;
 
 SEC("fexit/XXX")
 int BPF_PROG(fexit_XXX)
 {
 	struct bpf_perf_event_value *diff_val, *accum_val;
 	__u32 filter_key, zero = 0;
-	__u32 *accum_key;
+	__u32 accum_key;
+	struct bperf_filter_value *fval;
 
 	if (!enabled)
 		return 0;
 
 	switch (type) {
 	case BPERF_FILTER_GLOBAL:
-		accum_key = &zero;
+		accum_key = zero;
 		goto do_add;
 	case BPERF_FILTER_CPU:
 		filter_key = bpf_get_smp_processor_id();
@@ -49,22 +55,34 @@ int BPF_PROG(fexit_XXX)
 		filter_key = bpf_get_current_pid_tgid() & 0xffffffff;
 		break;
 	case BPERF_FILTER_TGID:
-		filter_key = bpf_get_current_pid_tgid() >> 32;
+		/* Use pid as the filter_key to exclude new task counts
+		 * when inherit is disabled. Don't worry about the existing
+		 * children in TGID losing their counts, bpf_counter has
+		 * already added them to the filter map via perf_thread_map
+		 * before this bpf prog runs.
+		 */
+		filter_key = inherit ?
+			     bpf_get_current_pid_tgid() >> 32 :
+			     bpf_get_current_pid_tgid() & 0xffffffff;
 		break;
 	default:
 		return 0;
 	}
 
-	accum_key = bpf_map_lookup_elem(&filter, &filter_key);
-	if (!accum_key)
+	fval = bpf_map_lookup_elem(&filter, &filter_key);
+	if (!fval)
 		return 0;
 
+	accum_key = fval->accum_key;
+	if (fval->exited)
+		bpf_map_delete_elem(&filter, &filter_key);
+
 do_add:
 	diff_val = bpf_map_lookup_elem(&diff_readings, &zero);
 	if (!diff_val)
 		return 0;
 
-	accum_val = bpf_map_lookup_elem(&accum_readings, accum_key);
+	accum_val = bpf_map_lookup_elem(&accum_readings, &accum_key);
 	if (!accum_val)
 		return 0;
 
@@ -75,4 +93,70 @@ do_add:
 	return 0;
 }
 
+/* The program is only used for PID or TGID filter types. */
+SEC("tp_btf/task_newtask")
+int BPF_PROG(on_newtask, struct task_struct *task, __u64 clone_flags)
+{
+	__u32 parent_key, child_key;
+	struct bperf_filter_value *parent_fval;
+	struct bperf_filter_value child_fval = { 0 };
+
+	if (!enabled)
+		return 0;
+
+	switch (type) {
+	case BPERF_FILTER_PID:
+		parent_key = bpf_get_current_pid_tgid() & 0xffffffff;
+		child_key = task->pid;
+		break;
+	case BPERF_FILTER_TGID:
+		parent_key = bpf_get_current_pid_tgid() >> 32;
+		child_key = task->tgid;
+		if (child_key == parent_key)
+			return 0;
+		break;
+	default:
+		return 0;
+	}
+
+	/* Check if the current task is one of the target tasks to be counted */
+	parent_fval = bpf_map_lookup_elem(&filter, &parent_key);
+	if (!parent_fval)
+		return 0;
+
+	/* Start counting for the new task by adding it into filter map,
+	 * inherit the accum key of its parent task so that they can be
+	 * counted together.
+	 */
+	child_fval.accum_key = parent_fval->accum_key;
+	child_fval.exited = 0;
+	bpf_map_update_elem(&filter, &child_key, &child_fval, BPF_NOEXIST);
+
+	return 0;
+}
+
+/* The program is only used for PID or TGID filter types. */
+SEC("tp_btf/sched_process_exit")
+int BPF_PROG(on_exittask, struct task_struct *task)
+{
+	__u32 pid;
+	struct bperf_filter_value *fval;
+
+	if (!enabled)
+		return 0;
+
+	/* Stop counting for this task by removing it from filter map.
+	 * For TGID type, if the pid can be found in the map, it means that
+	 * this pid belongs to the leader task. After the task exits, the
+	 * tgid of its child tasks (if any) will be 1, so the pid can be
+	 * safely removed.
+	 */
+	pid = task->pid;
+	fval = bpf_map_lookup_elem(&filter, &pid);
+	if (fval)
+		fval->exited = 1;
+
+	return 0;
+}
+
 char LICENSE[] SEC("license") = "Dual BSD/GPL";