simpleperf: update --print-hw-counter to check each cpu.

Little/Big cores may have different numbers of CPU PMU hardware counters.
So check them separately.

Bug: 243479304
Bug: 243034569
Test: run simpleperf_unit_test
Change-Id: I33527d69bc935b8523482feaf98378c5f6641762
(cherry picked from commit 6b771a2e12740f4bbcef515d772e62b364d87858)

3 files changed, 33 insertions, 8 deletions

diff --git a/simpleperf/cmd_stat.cpp b/simpleperf/cmd_stat.cpp
index 40a6bb0b..c46bd548 100644
--- a/simpleperf/cmd_stat.cpp
+++ b/simpleperf/cmd_stat.cpp
@@ -739,23 +739,23 @@ bool StatCommand::ParseOptions(const std::vector<std::string>& args,
   return true;
 }
 
-bool StatCommand::PrintHardwareCounters() {
+std::optional<size_t> GetHardwareCountersOnCpu(int cpu) {
   size_t available_counters = 0;
   const EventType* event = FindEventTypeByName("cpu-cycles", true);
   if (event == nullptr) {
-    return false;
+    return std::nullopt;
   }
   perf_event_attr attr = CreateDefaultPerfEventAttr(*event);
   while (true) {
     auto workload = Workload::CreateWorkload({"sleep", "0.1"});
-    if (!workload) {
-      return false;
+    if (!workload || !workload->SetCpuAffinity(cpu)) {
+      return std::nullopt;
     }
     std::vector<std::unique_ptr<EventFd>> event_fds;
     for (size_t i = 0; i <= available_counters; i++) {
       EventFd* group_event_fd = event_fds.empty() ? nullptr : event_fds[0].get();
-      auto event_fd =
-          EventFd::OpenEventFile(attr, workload->GetPid(), -1, group_event_fd, "cpu-cycles", false);
+      auto event_fd = EventFd::OpenEventFile(attr, workload->GetPid(), cpu, group_event_fd,
+                                             "cpu-cycles", false);
       if (!event_fd) {
         break;
       }
@@ -771,7 +771,7 @@ bool StatCommand::PrintHardwareCounters() {
     for (auto& event_fd : event_fds) {
       PerfCounter counter;
       if (!event_fd->ReadCounter(&counter)) {
-        return false;
+        return std::nullopt;
       }
       if (counter.time_enabled == 0 || counter.time_enabled > counter.time_running) {
         always_running = false;
@@ -783,7 +783,18 @@ bool StatCommand::PrintHardwareCounters() {
     }
     available_counters++;
   }
-  printf("There are %zu CPU PMU hardware counters available on this device.\n", available_counters);
+  return available_counters;
+}
+
+bool StatCommand::PrintHardwareCounters() {
+  for (int cpu : GetOnlineCpus()) {
+    std::optional<size_t> counters = GetHardwareCountersOnCpu(cpu);
+    if (!counters) {
+      LOG(ERROR) << "failed to get CPU PMU hardware counters on cpu " << cpu;
+      return false;
+    }
+    printf("There are %zu CPU PMU hardware counters available on cpu %d.\n", counters.value(), cpu);
+  }
   return true;
 }
 
diff --git a/simpleperf/workload.cpp b/simpleperf/workload.cpp
index 4bb7c540..142ed713 100644
--- a/simpleperf/workload.cpp
+++ b/simpleperf/workload.cpp
@@ -18,6 +18,7 @@
 
 #include <errno.h>
 #include <fcntl.h>
+#include <sched.h>
 #include <sys/prctl.h>
 #include <sys/wait.h>
 #include <unistd.h>
@@ -160,6 +161,18 @@ void Workload::ChildProcessFn(int start_signal_fd, int exec_child_fd) {
   }
 }
 
+bool Workload::SetCpuAffinity(int cpu) {
+  CHECK_EQ(work_state_, NotYetStartNewProcess);
+  cpu_set_t mask;
+  CPU_ZERO(&mask);
+  CPU_SET(cpu, &mask);
+  if (sched_setaffinity(GetPid(), sizeof(mask), &mask) != 0) {
+    PLOG(ERROR) << "sched_setaffinity failed";
+    return false;
+  }
+  return true;
+}
+
 bool Workload::Start() {
   CHECK_EQ(work_state_, NotYetStartNewProcess);
   char start_signal = 1;
diff --git a/simpleperf/workload.h b/simpleperf/workload.h
index 2400ac8f..86b28924 100644
--- a/simpleperf/workload.h
+++ b/simpleperf/workload.h
@@ -43,6 +43,7 @@ class Workload {
 
   ~Workload();
 
+  bool SetCpuAffinity(int cpu);
   bool Start();
   bool IsStarted() { return work_state_ == Started; }
   pid_t GetPid() { return work_pid_; }