perf intel-pt: Add support for PERF_RECORD_SWITCH
Add support for selecting and processing PERF_RECORD_SWITCH events for
use by Intel PT. If they are available, they will be used in preference
to sched_switch events.
This enables an unprivileged user to trace multi-threaded or
multi-process workloads with any level of perf_event_paranoid. However
it depends on kernel support for PERF_RECORD_SWITCH.
Without this patch, tracing a multi-threaded workload will decode
without error but all the data will be attributed to the main thread.
Without this patch, tracing a multi-process workload will result in
decoder errors because the decoder will not know which executable is
executing.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lkml.kernel.org/r/1439458857-30636-3-git-send-email-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
diff --git a/tools/perf/arch/x86/util/intel-pt.c b/tools/perf/arch/x86/util/intel-pt.c
index 2ca10d7..b02af06 100644
--- a/tools/perf/arch/x86/util/intel-pt.c
+++ b/tools/perf/arch/x86/util/intel-pt.c
@@ -624,13 +624,49 @@
* threads.
*/
if (have_timing_info && !cpu_map__empty(cpus)) {
- err = intel_pt_track_switches(evlist);
- if (err == -EPERM)
- pr_debug2("Unable to select sched:sched_switch\n");
- else if (err)
- return err;
- else
- ptr->have_sched_switch = 1;
+ if (perf_can_record_switch_events()) {
+ bool cpu_wide = !target__none(&opts->target) &&
+ !target__has_task(&opts->target);
+
+ if (!cpu_wide && perf_can_record_cpu_wide()) {
+ struct perf_evsel *switch_evsel;
+
+ err = parse_events(evlist, "dummy:u", NULL);
+ if (err)
+ return err;
+
+ switch_evsel = perf_evlist__last(evlist);
+
+ switch_evsel->attr.freq = 0;
+ switch_evsel->attr.sample_period = 1;
+ switch_evsel->attr.context_switch = 1;
+
+ switch_evsel->system_wide = true;
+ switch_evsel->no_aux_samples = true;
+ switch_evsel->immediate = true;
+
+ perf_evsel__set_sample_bit(switch_evsel, TID);
+ perf_evsel__set_sample_bit(switch_evsel, TIME);
+ perf_evsel__set_sample_bit(switch_evsel, CPU);
+
+ opts->record_switch_events = false;
+ ptr->have_sched_switch = 3;
+ } else {
+ opts->record_switch_events = true;
+ if (cpu_wide)
+ ptr->have_sched_switch = 3;
+ else
+ ptr->have_sched_switch = 2;
+ }
+ } else {
+ err = intel_pt_track_switches(evlist);
+ if (err == -EPERM)
+ pr_debug2("Unable to select sched:sched_switch\n");
+ else if (err)
+ return err;
+ else
+ ptr->have_sched_switch = 1;
+ }
}
if (intel_pt_evsel) {
@@ -663,8 +699,11 @@
tracking_evsel->attr.sample_period = 1;
/* In per-cpu case, always need the time of mmap events etc */
- if (!cpu_map__empty(cpus))
+ if (!cpu_map__empty(cpus)) {
perf_evsel__set_sample_bit(tracking_evsel, TIME);
+ /* And the CPU for switch events */
+ perf_evsel__set_sample_bit(tracking_evsel, CPU);
+ }
}
/*
diff --git a/tools/perf/util/intel-pt.c b/tools/perf/util/intel-pt.c
index bb41c20..2968b37 100644
--- a/tools/perf/util/intel-pt.c
+++ b/tools/perf/util/intel-pt.c
@@ -1145,11 +1145,13 @@
return 0;
}
-static u64 intel_pt_switch_ip(struct machine *machine, u64 *ptss_ip)
+static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
{
+ struct machine *machine = pt->machine;
struct map *map;
struct symbol *sym, *start;
u64 ip, switch_ip = 0;
+ const char *ptss;
if (ptss_ip)
*ptss_ip = 0;
@@ -1177,8 +1179,13 @@
if (!switch_ip || !ptss_ip)
return 0;
+ if (pt->have_sched_switch == 1)
+ ptss = "perf_trace_sched_switch";
+ else
+ ptss = "__perf_event_task_sched_out";
+
for (sym = start; sym; sym = dso__next_symbol(sym)) {
- if (!strcmp(sym->name, "perf_trace_sched_switch")) {
+ if (!strcmp(sym->name, ptss)) {
ip = map->unmap_ip(map, sym->start);
if (ip >= map->start && ip < map->end) {
*ptss_ip = ip;
@@ -1198,11 +1205,11 @@
if (!pt->kernel_start) {
pt->kernel_start = machine__kernel_start(pt->machine);
- if (pt->per_cpu_mmaps && pt->have_sched_switch &&
+ if (pt->per_cpu_mmaps &&
+ (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
!pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
!pt->sampling_mode) {
- pt->switch_ip = intel_pt_switch_ip(pt->machine,
- &pt->ptss_ip);
+ pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
if (pt->switch_ip) {
intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
pt->switch_ip, pt->ptss_ip);
@@ -1387,31 +1394,18 @@
return NULL;
}
-static int intel_pt_process_switch(struct intel_pt *pt,
- struct perf_sample *sample)
+static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
+ u64 timestamp)
{
struct intel_pt_queue *ptq;
- struct perf_evsel *evsel;
- pid_t tid;
- int cpu, err;
-
- evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
- if (evsel != pt->switch_evsel)
- return 0;
-
- tid = perf_evsel__intval(evsel, sample, "next_pid");
- cpu = sample->cpu;
-
- intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
- cpu, tid, sample->time, perf_time_to_tsc(sample->time,
- &pt->tc));
+ int err;
if (!pt->sync_switch)
- goto out;
+ return 1;
ptq = intel_pt_cpu_to_ptq(pt, cpu);
if (!ptq)
- goto out;
+ return 1;
switch (ptq->switch_state) {
case INTEL_PT_SS_NOT_TRACING:
@@ -1424,7 +1418,7 @@
return 0;
case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
if (!ptq->on_heap) {
- ptq->timestamp = perf_time_to_tsc(sample->time,
+ ptq->timestamp = perf_time_to_tsc(timestamp,
&pt->tc);
err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
ptq->timestamp);
@@ -1441,10 +1435,76 @@
default:
break;
}
-out:
+
+ return 1;
+}
+
+static int intel_pt_process_switch(struct intel_pt *pt,
+ struct perf_sample *sample)
+{
+ struct perf_evsel *evsel;
+ pid_t tid;
+ int cpu, ret;
+
+ evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
+ if (evsel != pt->switch_evsel)
+ return 0;
+
+ tid = perf_evsel__intval(evsel, sample, "next_pid");
+ cpu = sample->cpu;
+
+ intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
+ cpu, tid, sample->time, perf_time_to_tsc(sample->time,
+ &pt->tc));
+
+ ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
+ if (ret <= 0)
+ return ret;
+
return machine__set_current_tid(pt->machine, cpu, -1, tid);
}
+static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
+ struct perf_sample *sample)
+{
+ bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
+ pid_t pid, tid;
+ int cpu, ret;
+
+ cpu = sample->cpu;
+
+ if (pt->have_sched_switch == 3) {
+ if (!out)
+ return 0;
+ if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
+ pr_err("Expecting CPU-wide context switch event\n");
+ return -EINVAL;
+ }
+ pid = event->context_switch.next_prev_pid;
+ tid = event->context_switch.next_prev_tid;
+ } else {
+ if (out)
+ return 0;
+ pid = sample->pid;
+ tid = sample->tid;
+ }
+
+ if (tid == -1) {
+ pr_err("context_switch event has no tid\n");
+ return -EINVAL;
+ }
+
+ intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
+ cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
+ &pt->tc));
+
+ ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
+ if (ret <= 0)
+ return ret;
+
+ return machine__set_current_tid(pt->machine, cpu, pid, tid);
+}
+
static int intel_pt_process_itrace_start(struct intel_pt *pt,
union perf_event *event,
struct perf_sample *sample)
@@ -1515,6 +1575,9 @@
err = intel_pt_process_switch(pt, sample);
else if (event->header.type == PERF_RECORD_ITRACE_START)
err = intel_pt_process_itrace_start(pt, event, sample);
+ else if (event->header.type == PERF_RECORD_SWITCH ||
+ event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
+ err = intel_pt_context_switch(pt, event, sample);
intel_pt_log("event %s (%u): cpu %d time %"PRIu64" tsc %#"PRIx64"\n",
perf_event__name(event->header.type), event->header.type,
@@ -1777,6 +1840,18 @@
return NULL;
}
+static bool intel_pt_find_switch(struct perf_evlist *evlist)
+{
+ struct perf_evsel *evsel;
+
+ evlist__for_each(evlist, evsel) {
+ if (evsel->attr.context_switch)
+ return true;
+ }
+
+ return false;
+}
+
static const char * const intel_pt_info_fmts[] = {
[INTEL_PT_PMU_TYPE] = " PMU Type %"PRId64"\n",
[INTEL_PT_TIME_SHIFT] = " Time Shift %"PRIu64"\n",
@@ -1888,6 +1963,10 @@
pr_err("%s: missing sched_switch event\n", __func__);
goto err_delete_thread;
}
+ } else if (pt->have_sched_switch == 2 &&
+ !intel_pt_find_switch(session->evlist)) {
+ pr_err("%s: missing context_switch attribute flag\n", __func__);
+ goto err_delete_thread;
}
if (session->itrace_synth_opts && session->itrace_synth_opts->set) {