async-profiler#1044: Fall back to ctimer for CPU profiling when perf_…

…events are unavailable
SAP · Nov 4, 2024 · 6c32ce9 · 6c32ce9
1 parent adecac7
commit 6c32ce9
Show file tree

Hide file tree

Showing 16 changed files with 111 additions and 53 deletions.
diff --git a/src/allocTracer.h b/src/allocTracer.h
@@ -26,6 +26,10 @@ class AllocTracer : public Engine {
                                  uintptr_t total_size, uintptr_t instance_size);
 
   public:
+    const char* type() {
+        return "alloc_tracer";
+    }
+
     const char* title() {
         return "Allocation profile";
     }

diff --git a/src/arguments.cpp b/src/arguments.cpp
@@ -89,7 +89,6 @@ static const Multiplier UNIVERSAL[] = {{'n', 1}, {'u', 1000}, {'m', 1000000}, {'
 //     cstack=MODE      - how to collect C stack frames in addition to Java stack
 //                        MODE is 'fp', 'dwarf', 'lbr', 'vm' or 'no'
 //     clock=SOURCE     - clock source for JFR timestamps: 'tsc' or 'monotonic'
-//     allkernel        - include only kernel-mode events
 //     alluser          - include only user-mode events
 //     fdtransfer       - use fdtransfer to pass fds to the profiler
 //     simple           - simple class names instead of FQN
@@ -339,11 +338,8 @@ Error Arguments::parse(const char* args) {
             CASE("nobatch")
                 _nobatch = true;
 
-            CASE("allkernel")
-                _ring = RING_KERNEL;
-
             CASE("alluser")
-                _ring = RING_USER;
+                _alluser = true;
 
             CASE("cstack")
                 if (value != NULL) {

diff --git a/src/arguments.h b/src/arguments.h
@@ -41,12 +41,6 @@ enum SHORT_ENUM Counter {
     COUNTER_TOTAL
 };
 
-enum SHORT_ENUM Ring {
-    RING_ANY,
-    RING_KERNEL,
-    RING_USER
-};
-
 enum Style {
     STYLE_SIMPLE       = 0x1,
     STYLE_DOTTED       = 0x2,
@@ -159,7 +153,6 @@ class Arguments {
   public:
     Action _action;
     Counter _counter;
-    Ring _ring;
     const char* _event;
     int _timeout;
     long _interval;
@@ -183,6 +176,7 @@ class Arguments {
     bool _sched;
     bool _live;
     bool _nobatch;
+    bool _alluser;
     bool _fdtransfer;
     const char* _fdtransfer_path;
     int _style;
@@ -209,7 +203,6 @@ class Arguments {
         _shared(false),
         _action(ACTION_NONE),
         _counter(COUNTER_SAMPLES),
-        _ring(RING_ANY),
         _event(NULL),
         _timeout(0),
         _interval(0),
@@ -233,6 +226,7 @@ class Arguments {
         _sched(false),
         _live(false),
         _nobatch(false),
+        _alluser(false),
         _fdtransfer(false),
         _fdtransfer_path(NULL),
         _style(0),

diff --git a/src/ctimer.h b/src/ctimer.h
@@ -19,6 +19,10 @@ class CTimer : public CpuEngine {
     void destroyForThread(int tid);
 
   public:
+    const char* type() {
+        return "ctimer";
+    }
+
     Error check(Arguments& args);
     Error start(Arguments& args);
     void stop();

diff --git a/src/engine.h b/src/engine.h
@@ -34,6 +34,10 @@ class Engine {
     }
 
   public:
+    virtual const char* type() {
+        return "noop";
+    }
+
     virtual const char* title() {
         return "Flame Graph";
     }

diff --git a/src/flightRecorder.cpp b/src/flightRecorder.cpp
@@ -56,7 +56,6 @@ static jmethodID _start_method;
 static jmethodID _stop_method;
 static jmethodID _box_method;
 
-static const char* const SETTING_RING[] = {NULL, "kernel", "user"};
 static const char* const SETTING_CSTACK[] = {NULL, "no", "fp", "dwarf", "lbr", "vm"};
 static const char* const SETTING_CLOCK[] = {NULL, "tsc", "monotonic"};
 
@@ -804,7 +803,7 @@ class Recording {
 
     void writeSettings(Buffer* buf, Arguments& args) {
         writeStringSetting(buf, T_ACTIVE_RECORDING, "version", PROFILER_VERSION);
-        writeStringSetting(buf, T_ACTIVE_RECORDING, "ring", SETTING_RING[args._ring]);
+        writeStringSetting(buf, T_ACTIVE_RECORDING, "engine", Profiler::instance()->_engine->type());
         writeStringSetting(buf, T_ACTIVE_RECORDING, "cstack", SETTING_CSTACK[args._cstack]);
         writeStringSetting(buf, T_ACTIVE_RECORDING, "clock", SETTING_CLOCK[args._clock]);
         writeStringSetting(buf, T_ACTIVE_RECORDING, "event", args._event);
@@ -824,6 +823,7 @@ class Recording {
         writeBoolSetting(buf, T_EXECUTION_SAMPLE, "enabled", args._event != NULL);
         if (args._event != NULL) {
             writeIntSetting(buf, T_EXECUTION_SAMPLE, "interval", args._interval);
+            writeBoolSetting(buf, T_EXECUTION_SAMPLE, "alluser", args._alluser);
         }
         if (args._wall >= 0) {
             writeIntSetting(buf, T_EXECUTION_SAMPLE, "wall", args._wall);

diff --git a/src/instrument.h b/src/instrument.h
@@ -19,6 +19,10 @@ class Instrument : public Engine {
     static volatile bool _running;
 
   public:
+    const char* type() {
+        return "instrument";
+    }
+
     const char* title() {
         return "Java method profile";
     }

diff --git a/src/itimer.h b/src/itimer.h
@@ -11,6 +11,10 @@
 
 class ITimer : public CpuEngine {
   public:
+    const char* type() {
+        return "itimer";
+    }
+
     Error check(Arguments& args);
     Error start(Arguments& args);
     void stop();

diff --git a/src/j9ObjectSampler.h b/src/j9ObjectSampler.h
@@ -11,6 +11,10 @@
 
 class J9ObjectSampler : public ObjectSampler {
   public:
+    const char* type() {
+        return "j9_object_sampler";
+    }
+
     Error check(Arguments& args);
     Error start(Arguments& args);
     void stop();

diff --git a/src/j9WallClock.h b/src/j9WallClock.h
@@ -26,6 +26,10 @@ class J9WallClock : public Engine {
     void timerLoop();
 
   public:
+    const char* type() {
+        return "j9_wall";
+    }
+
     const char* title() {
         return "Wall clock profile";
     }

diff --git a/src/lockTracer.h b/src/lockTracer.h
@@ -47,6 +47,10 @@ class LockTracer : public Engine {
                                     const char* lock_name, jobject lock, jlong timeout);
 
   public:
+    const char* type() {
+        return "lock_tracer";
+    }
+
     const char* title() {
         return "Lock profile";
     }

diff --git a/src/objectSampler.h b/src/objectSampler.h
@@ -25,6 +25,10 @@ class ObjectSampler : public Engine {
                                  jobject object, jclass object_klass, jlong size);
 
   public:
+    const char* type() {
+        return "object_sampler";
+    }
+
     const char* title() {
         return "Allocation profile";
     }

diff --git a/src/perfEvents.h b/src/perfEvents.h
@@ -20,8 +20,8 @@ class PerfEvents : public CpuEngine {
     static int _max_events;
     static PerfEvent* _events;
     static PerfEventType* _event_type;
-    static Ring _ring;
-    static bool _use_mmap_page;
+    static bool _alluser;
+    static bool _kernel_stack;
 
     static u64 readCounter(siginfo_t* siginfo, void* ucontext);
     static void signalHandler(int signo, siginfo_t* siginfo, void* ucontext);
@@ -35,6 +35,10 @@ class PerfEvents : public CpuEngine {
     Error start(Arguments& args);
     void stop();
 
+    const char* type() {
+        return "perf";
+    }
+
     const char* title();
     const char* units();
 

diff --git a/src/perfEvents_linux.cpp b/src/perfEvents_linux.cpp
@@ -166,6 +166,7 @@ struct PerfEventType {
     int counter_arg;
 
     enum {
+        IDX_CPU = 0,
         IDX_PREDEFINED = 12,
         IDX_RAW,
         IDX_PMU,
@@ -355,6 +356,11 @@ struct PerfEventType {
     }
 
     static PerfEventType* forName(const char* name) {
+        // "cpu" is an alias for "cpu-clock"
+        if (strcmp(name, EVENT_CPU) == 0) {
+            return &AVAILABLE_EVENTS[IDX_CPU];
+        }
+
         // Look through the table of predefined perf events
         for (int i = 0; i <= IDX_PREDEFINED; i++) {
             if (strcmp(name, AVAILABLE_EVENTS[i].name) == 0) {
@@ -429,7 +435,7 @@ struct PerfEventType {
 #endif
 
 PerfEventType PerfEventType::AVAILABLE_EVENTS[] = {
-    {"cpu",          DEFAULT_INTERVAL, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_CPU_CLOCK},
+    {"cpu-clock",    DEFAULT_INTERVAL, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_CPU_CLOCK},
     {"page-faults",                 1, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_PAGE_FAULTS},
     {"context-switches",            2, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_CONTEXT_SWITCHES},
 
@@ -445,6 +451,8 @@ PerfEventType PerfEventType::AVAILABLE_EVENTS[] = {
     {"LLC-load-misses",          1000, PERF_TYPE_HW_CACHE, LOAD_MISS(PERF_COUNT_HW_CACHE_LL)},
     {"dTLB-load-misses",         1000, PERF_TYPE_HW_CACHE, LOAD_MISS(PERF_COUNT_HW_CACHE_DTLB)},
 
+    /* End of IDX_PREDEFINED events */
+
     {"rNNN",                     1000, PERF_TYPE_RAW, 0}, /* IDX_RAW */
     {"pmu/event-descriptor/",    1000, PERF_TYPE_RAW, 0}, /* IDX_PMU */
 
@@ -510,8 +518,8 @@ class PerfEvent : public SpinLock {
 int PerfEvents::_max_events = 0;
 PerfEvent* PerfEvents::_events = NULL;
 PerfEventType* PerfEvents::_event_type = NULL;
-Ring PerfEvents::_ring;
-bool PerfEvents::_use_mmap_page;
+bool PerfEvents::_alluser;
+bool PerfEvents::_kernel_stack;
 
 int PerfEvents::createForThread(int tid) {
     if (tid >= _max_events) {
@@ -548,10 +556,12 @@ int PerfEvents::createForThread(int tid) {
     attr.disabled = 1;
     attr.wakeup_events = 1;
 
-    if (_ring == RING_USER) {
+    if (_alluser) {
         attr.exclude_kernel = 1;
-    } else if (_ring == RING_KERNEL) {
-        attr.exclude_user = 1;
+    }
+
+    if (!_kernel_stack) {
+        attr.exclude_callchain_kernel = 1;
     }
 
     if (_cstack >= CSTACK_FP) {
@@ -586,10 +596,13 @@ int PerfEvents::createForThread(int tid) {
         return err;
     }
 
-    void* page = _use_mmap_page ? mmap(NULL, 2 * OS::page_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0) : NULL;
-    if (page == MAP_FAILED) {
-        Log::warn("perf_event mmap failed: %s", strerror(errno));
-        page = NULL;
+    void* page = NULL;
+    if (_kernel_stack || _cstack == CSTACK_DEFAULT || _cstack == CSTACK_LBR) {
+        page = mmap(NULL, 2 * OS::page_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+        if (page == MAP_FAILED) {
+            Log::warn("perf_event mmap failed: %s", strerror(errno));
+            page = NULL;
+        }
     }
 
     _events[tid].reset();
@@ -693,7 +706,7 @@ void PerfEvents::signalHandlerJ9(int signo, siginfo_t* siginfo, void* ucontext)
 }
 
 const char* PerfEvents::title() {
-    if (_event_type == NULL || _event_type->name == EVENT_CPU) {
+    if (_event_type == NULL || strcmp(_event_type->name, "cpu-clock") == 0) {
         return "CPU profile";
     } else if (_event_type->type == PERF_TYPE_SOFTWARE || _event_type->type == PERF_TYPE_HARDWARE || _event_type->type == PERF_TYPE_HW_CACHE) {
         return _event_type->name;
@@ -703,7 +716,7 @@ const char* PerfEvents::title() {
 }
 
 const char* PerfEvents::units() {
-    return _event_type == NULL || _event_type->name == EVENT_CPU ? "ns" : "total";
+    return _event_type == NULL || strcmp(_event_type->name, "cpu-clock") == 0 ? "ns" : "total";
 }
 
 Error PerfEvents::check(Arguments& args) {
@@ -734,17 +747,8 @@ Error PerfEvents::check(Arguments& args) {
     attr.sample_type = PERF_SAMPLE_CALLCHAIN;
     attr.disabled = 1;
 
-    if (args._ring == RING_USER) {
+    if (args._alluser) {
         attr.exclude_kernel = 1;
-    } else if (args._ring == RING_KERNEL) {
-        attr.exclude_user = 1;
-    } else if (!Symbols::haveKernelSymbols()) {
-        Profiler::instance()->updateSymbols(true);
-        attr.exclude_kernel = Symbols::haveKernelSymbols() ? 0 : 1;
-    }
-
-    if (args._cstack >= CSTACK_FP) {
-        attr.exclude_callchain_user = 1;
     }
 
 #ifdef PERF_ATTR_SIZE_VER5
@@ -783,14 +787,14 @@ Error PerfEvents::start(Arguments& args) {
     _cstack = args._cstack;
     _signal = args._signal == 0 ? OS::getProfilingSignal(0) : args._signal & 0xff;
 
-    _ring = args._ring;
-    if (_ring != RING_USER && !Symbols::haveKernelSymbols()) {
+    _alluser = args._alluser;
+    _kernel_stack = !_alluser && _cstack != CSTACK_NO;
+    if (_kernel_stack && !Symbols::haveKernelSymbols()) {
         Log::warn("Kernel symbols are unavailable due to restrictions. Try\n"
                   "  sysctl kernel.perf_event_paranoid=1\n"
                   "  sysctl kernel.kptr_restrict=0");
-        _ring = RING_USER;
+        _kernel_stack = false;
     }
-    _use_mmap_page = _cstack != CSTACK_NO && (_ring != RING_USER || _cstack == CSTACK_DEFAULT || _cstack == CSTACK_LBR);
 
     adjustFDLimit();
 
@@ -820,7 +824,7 @@ Error PerfEvents::start(Arguments& args) {
     if (err) {
         stop();
         if (err == EACCES || err == EPERM) {
-            return Error("No access to perf events. Try --fdtransfer or --all-user option or 'sysctl kernel.perf_event_paranoid=1'");
+            return Error("Perf events unavailable. Try --fdtransfer or --all-user option or 'sysctl kernel.perf_event_paranoid=1'");
         } else if (isResourceLimit(err)) {
             return Error("Perf events resource limit. Check 'ulimit -n'");
         } else {
@@ -938,10 +942,21 @@ void PerfEvents::resetBuffer(int tid) {
 }
 
 bool PerfEvents::supported() {
-    // The official way of knowing if perf_event_open() support is enabled
-    // is checking for the existence of the file /proc/sys/kernel/perf_event_paranoid
-    struct stat statbuf;
-    return stat("/proc/sys/kernel/perf_event_paranoid", &statbuf) == 0;
+    struct perf_event_attr attr = {0};
+    attr.size = sizeof(attr);
+    attr.type = PERF_TYPE_SOFTWARE;
+    attr.config = PERF_COUNT_SW_CPU_CLOCK;
+    attr.sample_period = 1000000000;
+    attr.sample_type = PERF_SAMPLE_CALLCHAIN;
+    attr.disabled = 1;
+
+    int fd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
+    if (fd == -1) {
+        return false;
+    }
+
+    close(fd);
+    return true;
 }
 
 const char* PerfEvents::getEventName(int event_id) {