vernier 0.1.1 → 0.2.0

data/ext/vernier/vernier.cc

@@ -1,27 +1,276 @@
 #include <iostream>
+#include <iomanip>
 #include <vector>
 #include <memory>
 #include <algorithm>
 #include <sstream>
 #include <unordered_map>
 #include <unordered_set>
+#include <cassert>
+#include <atomic>
+#include <mutex>
+#include <optional>
+
+#include <sys/time.h>
+#include <signal.h>
+#ifdef __APPLE__
+#include <dispatch/dispatch.h>
+#else
+#include <semaphore.h>
+#endif
 
 #include "vernier.hh"
 #include "stack.hh"
+
+#include "ruby/ruby.h"
 #include "ruby/debug.h"
+#include "ruby/thread.h"
 
-using namespace std;
+// GC event's we'll monitor during profiling
+#define RUBY_GC_PHASE_EVENTS \
+  RUBY_INTERNAL_EVENT_GC_START | \
+  RUBY_INTERNAL_EVENT_GC_END_MARK | \
+  RUBY_INTERNAL_EVENT_GC_END_SWEEP | \
+  RUBY_INTERNAL_EVENT_GC_ENTER | \
+  RUBY_INTERNAL_EVENT_GC_EXIT
+
+#define sym(name) ID2SYM(rb_intern_const(name))
+
+// HACK: This isn't public, but the objspace ext uses it
+extern "C" size_t rb_obj_memsize_of(VALUE);
 
-#define numberof(array) ((int)(sizeof(array) / sizeof((array)[0])))
+using namespace std;
 
 static VALUE rb_mVernier;
+static VALUE rb_cVernierResult;
+static VALUE rb_mVernierMarkerType;
+static VALUE rb_cVernierCollector;
+
+class TimeStamp {
+    static const uint64_t nanoseconds_per_second = 1000000000;
+    uint64_t value_ns;
+
+    TimeStamp(uint64_t value_ns) : value_ns(value_ns) {}
+
+    public:
+    TimeStamp() : value_ns(0) {}
+
+    static TimeStamp Now() {
+        struct timespec ts;
+        clock_gettime(CLOCK_MONOTONIC, &ts);
+        return TimeStamp(ts.tv_sec * nanoseconds_per_second + ts.tv_nsec);
+    }
+
+    static TimeStamp Zero() {
+        return TimeStamp(0);
+    }
+
+    static void Sleep(const TimeStamp &time) {
+        struct timespec ts = time.timespec();
+
+        int res;
+        do {
+            res = nanosleep(&ts, &ts);
+        } while (res && errno == EINTR);
+    }
+
+    static TimeStamp from_microseconds(uint64_t us) {
+        return TimeStamp(us * 1000);
+    }
+
+    static TimeStamp from_nanoseconds(uint64_t ns) {
+        return TimeStamp(ns);
+    }
+
+    TimeStamp operator-(const TimeStamp &other) const {
+        TimeStamp result = *this;
+        return result -= other;
+    }
+
+    TimeStamp &operator-=(const TimeStamp &other) {
+        if (value_ns > other.value_ns) {
+            value_ns = value_ns - other.value_ns;
+        } else {
+            // underflow
+            value_ns = 0;
+        }
+        return *this;
+    }
+
+    TimeStamp operator+(const TimeStamp &other) const {
+        TimeStamp result = *this;
+        return result += other;
+    }
+
+    TimeStamp &operator+=(const TimeStamp &other) {
+        uint64_t new_value = value_ns + other.value_ns;
+        value_ns = new_value;
+        return *this;
+    }
+
+    bool operator<(const TimeStamp &other) const {
+        return value_ns < other.value_ns;
+    }
+
+    bool operator<=(const TimeStamp &other) const {
+        return value_ns <= other.value_ns;
+    }
+
+    bool operator>(const TimeStamp &other) const {
+        return value_ns > other.value_ns;
+    }
+
+    bool operator>=(const TimeStamp &other) const {
+        return value_ns >= other.value_ns;
+    }
+
+    uint64_t nanoseconds() const {
+        return value_ns;
+    }
+
+    uint64_t microseconds() const {
+        return value_ns / 1000;
+    }
+
+    bool zero() const {
+        return value_ns == 0;
+    }
+
+    struct timespec timespec() const {
+        struct timespec ts;
+        ts.tv_sec = nanoseconds() / nanoseconds_per_second;
+        ts.tv_nsec = (nanoseconds() % nanoseconds_per_second);
+        return ts;
+    }
+};
+
+std::ostream& operator<<(std::ostream& os, const TimeStamp& info) {
+    os << info.nanoseconds() << "ns";
+    return os;
+}
 
-struct retained_collector {
-    int allocated_objects = 0;
-    int freed_objects = 0;
+// A basic semaphore built on sem_wait/sem_post
+// post() is guaranteed to be async-signal-safe
+class SamplerSemaphore {
+#ifdef __APPLE__
+    dispatch_semaphore_t sem;
+#else
+    sem_t sem;
+#endif
 
-    std::unordered_set<VALUE> unique_frames;
-    std::unordered_map<VALUE, std::unique_ptr<Stack>> object_frames;
+    public:
+
+    SamplerSemaphore(unsigned int value = 0) {
+#ifdef __APPLE__
+        sem = dispatch_semaphore_create(value);
+#else
+        sem_init(&sem, 0, value);
+#endif
+    };
+
+    ~SamplerSemaphore() {
+#ifdef __APPLE__
+        dispatch_release(sem);
+#else
+        sem_destroy(&sem);
+#endif
+    };
+
+    void wait() {
+#ifdef __APPLE__
+        dispatch_semaphore_wait(sem, DISPATCH_TIME_FOREVER);
+#else
+        int ret;
+        do {
+            ret = sem_wait(&sem);
+        } while (ret && errno == EINTR);
+        assert(ret == 0);
+#endif
+    }
+
+    void post() {
+#ifdef __APPLE__
+        dispatch_semaphore_signal(sem);
+#else
+        sem_post(&sem);
+#endif
+    }
+};
+
+struct RawSample {
+    constexpr static int MAX_LEN = 2048;
+    VALUE frames[MAX_LEN];
+    int lines[MAX_LEN];
+    int len;
+    bool gc;
+
+    RawSample() : len(0), gc(false) { }
+
+    int size() const {
+        return len;
+    }
+
+    Frame frame(int i) const {
+        const Frame frame = {frames[i], lines[i]};
+        return frame;
+    }
+
+    void sample() {
+        if (!ruby_native_thread_p()) {
+            clear();
+            return;
+        }
+
+        if (rb_during_gc()) {
+          gc = true;
+          len = 0;
+        } else {
+          gc = false;
+          len = rb_profile_frames(0, MAX_LEN, frames, lines);
+        }
+    }
+
+    void clear() {
+        len = 0;
+        gc = false;
+    }
+
+    bool empty() const {
+        return len == 0;
+    }
+};
+
+// Based very loosely on the design of Gecko's SigHandlerCoordinator
+// This is used for communication between the profiler thread and the signal
+// handlers in the observed thread.
+struct LiveSample {
+    RawSample sample;
+
+    SamplerSemaphore sem_complete;
+
+    // Wait for a sample to be collected by the signal handler on another thread
+    void wait() {
+        sem_complete.wait();
+    }
+
+    int size() const {
+        return sample.size();
+    }
+
+    Frame frame(int i) const {
+        return sample.frame(i);
+    }
+
+    // Called from a signal handler in the observed thread in order to take a
+    // sample and signal to the proifiler thread that the sample is ready.
+    //
+    // CRuby doesn't guarantee that rb_profile_frames can be used as
+    // async-signal-safe but in practice it seems to be.
+    // sem_post is safe in an async-signal-safe context.
+    void sample_current_thread() {
+        sample.sample();
+        sem_complete.post();
+    }
 };
 
 struct TraceArg {
@@ -42,149 +291,934 @@ struct TraceArg {
     }
 };
 
-static retained_collector _collector;
+struct FrameList {
+    std::unordered_map<std::string, int> string_to_idx;
+    std::vector<std::string> string_list;
 
-static VALUE tp_newobj;
-static VALUE tp_freeobj;
-static void
-newobj_i(VALUE tpval, void *data) {
-    retained_collector *collector = static_cast<retained_collector *>(data);
-    TraceArg tp(tpval);
-    collector->allocated_objects++;
+    int string_index(const std::string str) {
+        auto it = string_to_idx.find(str);
+        if (it == string_to_idx.end()) {
+            int idx = string_list.size();
+            string_list.push_back(str);
 
-    VALUE frames_buffer[2048];
-    int lines_buffer[2048];
-    int n = rb_profile_frames(0, 2048, frames_buffer, lines_buffer);
+            auto result = string_to_idx.insert({str, idx});
+            it = result.first;
+        }
 
-    for (int i = 0; i < n; i++) {
-        collector->unique_frames.insert(frames_buffer[i]);
+        return it->second;
     }
 
-    collector->object_frames.emplace(
-            tp.obj,
-            make_unique<Stack>(frames_buffer, lines_buffer, n)
-            );
-}
+    struct FrameWithInfo {
+        Frame frame;
+        FrameInfo info;
+    };
+
+    std::unordered_map<Frame, int> frame_to_idx;
+    std::vector<Frame> frame_list;
+    std::vector<FrameWithInfo> frame_with_info_list;
+    int frame_index(const Frame frame) {
+        auto it = frame_to_idx.find(frame);
+        if (it == frame_to_idx.end()) {
+            int idx = frame_list.size();
+            frame_list.push_back(frame);
+            auto result = frame_to_idx.insert({frame, idx});
+            it = result.first;
+        }
+        return it->second;
+    }
+
+    struct StackNode {
+        std::unordered_map<Frame, int> children;
+        Frame frame;
+        int parent;
+        int index;
+
+        StackNode(Frame frame, int index, int parent) : frame(frame), index(index), parent(parent) {}
+
+        // root
+        StackNode() : frame(Frame{0, 0}), index(-1), parent(-1) {}
+    };
+
+    StackNode root_stack_node;
+    vector<StackNode> stack_node_list;
+
+    int stack_index(const RawSample &stack) {
+        if (stack.empty()) {
+            throw std::runtime_error("empty stack");
+        }
+
+        StackNode *node = &root_stack_node;
+        for (int i = stack.size() - 1; i >= 0; i--) {
+            const Frame &frame = stack.frame(i);
+            node = next_stack_node(node, frame);
+        }
+        return node->index;
+    }
+
+    StackNode *next_stack_node(StackNode *node, const Frame &frame) {
+        int next_node_idx = node->children[frame];
+        if (next_node_idx == 0) {
+            // insert a new node
+            next_node_idx = stack_node_list.size();
+            node->children[frame] = next_node_idx;
+            stack_node_list.emplace_back(
+                    frame,
+                    next_node_idx,
+                    node->index
+                    );
+        }
+
+        return &stack_node_list[next_node_idx];
+    }
+
+    // Converts Frames from stacks other tables. "Symbolicates" the frames
+    // which allocates.
+    void finalize() {
+        for (const auto &stack_node : stack_node_list) {
+            frame_index(stack_node.frame);
+        }
+        for (const auto &frame : frame_list) {
+            frame_with_info_list.push_back(FrameWithInfo{frame, frame.info()});
+        }
+    }
+
+    void mark_frames() {
+        for (auto stack_node: stack_node_list) {
+            rb_gc_mark(stack_node.frame.frame);
+        }
+    }
+
+    void clear() {
+        string_list.clear();
+        frame_list.clear();
+        stack_node_list.clear();
+        frame_with_info_list.clear();
+
+        string_to_idx.clear();
+        frame_to_idx.clear();
+        root_stack_node.children.clear();
+    }
+
+    void write_result(VALUE result) {
+        FrameList &frame_list = *this;
+
+        VALUE stack_table = rb_hash_new();
+        rb_ivar_set(result, rb_intern("@stack_table"), stack_table);
+        VALUE stack_table_parent = rb_ary_new();
+        VALUE stack_table_frame = rb_ary_new();
+        rb_hash_aset(stack_table, sym("parent"), stack_table_parent);
+        rb_hash_aset(stack_table, sym("frame"), stack_table_frame);
+        for (const auto &stack : frame_list.stack_node_list) {
+            VALUE parent_val = stack.parent == -1 ? Qnil : INT2NUM(stack.parent);
+            rb_ary_push(stack_table_parent, parent_val);
+            rb_ary_push(stack_table_frame, INT2NUM(frame_list.frame_index(stack.frame)));
+        }
+
+        VALUE frame_table = rb_hash_new();
+        rb_ivar_set(result, rb_intern("@frame_table"), frame_table);
+        VALUE frame_table_func = rb_ary_new();
+        VALUE frame_table_line = rb_ary_new();
+        rb_hash_aset(frame_table, sym("func"), frame_table_func);
+        rb_hash_aset(frame_table, sym("line"), frame_table_line);
+        //for (const auto &frame : frame_list.frame_list) {
+        for (int i = 0; i < frame_list.frame_with_info_list.size(); i++) {
+            const auto &frame = frame_list.frame_with_info_list[i];
+            rb_ary_push(frame_table_func, INT2NUM(i));
+            rb_ary_push(frame_table_line, INT2NUM(frame.frame.line));
+        }
+
+        // TODO: dedup funcs before this step
+        VALUE func_table = rb_hash_new();
+        rb_ivar_set(result, rb_intern("@func_table"), func_table);
+        VALUE func_table_name = rb_ary_new();
+        VALUE func_table_filename = rb_ary_new();
+        VALUE func_table_first_line = rb_ary_new();
+        rb_hash_aset(func_table, sym("name"), func_table_name);
+        rb_hash_aset(func_table, sym("filename"), func_table_filename);
+        rb_hash_aset(func_table, sym("first_line"), func_table_first_line);
+        for (const auto &frame : frame_list.frame_with_info_list) {
+            const std::string label = frame.info.label;
+            const std::string filename = frame.info.file;
+            const int first_line = frame.info.first_lineno;
+
+            rb_ary_push(func_table_name, rb_str_new(label.c_str(), label.length()));
+            rb_ary_push(func_table_filename, rb_str_new(filename.c_str(), filename.length()));
+            rb_ary_push(func_table_first_line, INT2NUM(first_line));
+        }
+    }
+};
+
+class BaseCollector {
+    protected:
+
+    virtual void reset() {
+        frame_list.clear();
+    }
+
+    public:
+    bool running = false;
+    FrameList frame_list;
+
+    virtual bool start() {
+        if (running) {
+            return false;
+        } else {
+            running = true;
+            return true;
+        }
+    }
+
+    virtual VALUE stop() {
+        if (!running) {
+            rb_raise(rb_eRuntimeError, "collector not running");
+        }
+        running = false;
+
+        return Qnil;
+    }
+
+    virtual void sample() {
+        rb_raise(rb_eRuntimeError, "collector doesn't support manual sampling");
+    };
+
+    virtual void mark() {
+        frame_list.mark_frames();
+    };
+
+    virtual VALUE get_markers() {
+        return rb_ary_new();
+    };
+};
+
+class CustomCollector : public BaseCollector {
+    std::vector<int> samples;
+
+    void sample() {
+        RawSample sample;
+        sample.sample();
+        int stack_index = frame_list.stack_index(sample);
+
+        samples.push_back(stack_index);
+    }
+
+    VALUE stop() {
+        BaseCollector::stop();
+
+        frame_list.finalize();
+
+        VALUE result = build_collector_result();
+
+        reset();
+
+        return result;
+    }
+
+    VALUE build_collector_result() {
+        VALUE result = rb_obj_alloc(rb_cVernierResult);
+
+        VALUE samples = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@samples"), samples);
+        VALUE weights = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@weights"), weights);
+
+        for (auto& stack_index: this->samples) {
+            rb_ary_push(samples, INT2NUM(stack_index));
+            rb_ary_push(weights, INT2NUM(1));
+        }
+
+        frame_list.write_result(result);
+
+        return result;
+    }
+};
+
+class RetainedCollector : public BaseCollector {
+    void reset() {
+        object_frames.clear();
+        object_list.clear();
+
+        BaseCollector::reset();
+    }
+
+    void record(VALUE obj) {
+        RawSample sample;
+        sample.sample();
+        int stack_index = frame_list.stack_index(sample);
+
+        object_list.push_back(obj);
+        object_frames.emplace(obj, stack_index);
+    }
+
+    std::unordered_map<VALUE, int> object_frames;
+    std::vector<VALUE> object_list;
+
+    VALUE tp_newobj = Qnil;
+    VALUE tp_freeobj = Qnil;
+
+    static void newobj_i(VALUE tpval, void *data) {
+        RetainedCollector *collector = static_cast<RetainedCollector *>(data);
+        TraceArg tp(tpval);
+
+        collector->record(tp.obj);
+    }
+
+    static void freeobj_i(VALUE tpval, void *data) {
+        RetainedCollector *collector = static_cast<RetainedCollector *>(data);
+        TraceArg tp(tpval);
+
+        collector->object_frames.erase(tp.obj);
+    }
+
+    public:
+
+    bool start() {
+        if (!BaseCollector::start()) {
+            return false;
+        }
+
+        tp_newobj = rb_tracepoint_new(0, RUBY_INTERNAL_EVENT_NEWOBJ, newobj_i, this);
+        tp_freeobj = rb_tracepoint_new(0, RUBY_INTERNAL_EVENT_FREEOBJ, freeobj_i, this);
+
+        rb_tracepoint_enable(tp_newobj);
+        rb_tracepoint_enable(tp_freeobj);
+
+        return true;
+    }
+
+    VALUE stop() {
+        BaseCollector::stop();
+
+        // GC before we start turning stacks into strings
+        rb_gc();
+
+        // Stop tracking any more new objects, but we'll continue tracking free'd
+        // objects as we may be able to free some as we remove our own references
+        // to stack frames.
+        rb_tracepoint_disable(tp_newobj);
+        tp_newobj = Qnil;
+
+        frame_list.finalize();
+
+        // We should have collected info for all our frames, so no need to continue
+        // marking them
+        // FIXME: previously here we cleared the list of frames so we would stop
+        // marking them. Maybe now we should set a flag so that we stop marking them
+
+        // GC again
+        rb_gc();
+
+        rb_tracepoint_disable(tp_freeobj);
+        tp_freeobj = Qnil;
+
+        VALUE result = build_collector_result();
+
+        reset();
+
+        return result;
+    }
+
+    VALUE build_collector_result() {
+        RetainedCollector *collector = this;
+        FrameList &frame_list = collector->frame_list;
+
+        VALUE result = rb_obj_alloc(rb_cVernierResult);
+
+        VALUE samples = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@samples"), samples);
+        VALUE weights = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@weights"), weights);
+
+        for (auto& obj: collector->object_list) {
+            const auto search = collector->object_frames.find(obj);
+            if (search != collector->object_frames.end()) {
+                int stack_index = search->second;
+
+                rb_ary_push(samples, INT2NUM(stack_index));
+                rb_ary_push(weights, INT2NUM(rb_obj_memsize_of(obj)));
+            }
+        }
+
+        frame_list.write_result(result);
+
+        return result;
+    }
+
+    void mark() {
+        // We don't mark the objects, but we MUST mark the frames, otherwise they
+        // can be garbage collected.
+        // When we stop collection we will stringify the remaining frames, and then
+        // clear them from the set, allowing them to be removed from out output.
+        frame_list.mark_frames();
+
+        rb_gc_mark(tp_newobj);
+        rb_gc_mark(tp_freeobj);
+    }
+};
+
+typedef uint64_t native_thread_id_t;
+
+class Thread {
+    public:
+        static native_thread_id_t get_native_thread_id() {
+#ifdef __APPLE__
+            uint64_t thread_id;
+            int e = pthread_threadid_np(pthread_self(), &thread_id);
+            if (e != 0) rb_syserr_fail(e, "pthread_threadid_np");
+            return thread_id;
+#else
+            return gettid();
+#endif
+        }
+
+        enum State {
+            STARTED,
+            RUNNING,
+            SUSPENDED,
+            STOPPED
+        };
+
+        pthread_t pthread_id;
+        native_thread_id_t native_tid;
+        State state;
+
+        TimeStamp state_changed_at;
+        TimeStamp started_at;
+        TimeStamp stopped_at;
+
+        RawSample stack_on_suspend;
+
+        std::string name;
+
+        Thread(State state) : state(state) {
+            pthread_id = pthread_self();
+            native_tid = get_native_thread_id();
+            started_at = state_changed_at = TimeStamp::Now();
+        }
+
+        void set_state(State new_state) {
+            if (state == Thread::State::STOPPED) {
+                return;
+            }
+
+            auto now = TimeStamp::Now();
+
+            state = new_state;
+            state_changed_at = now;
+            if (new_state == State::STARTED) {
+                if (started_at.zero()) {
+                    started_at = now;
+                }
+            } else if (new_state == State::STOPPED) {
+                stopped_at = now;
+
+                capture_name();
+            }
+        }
+
+        bool running() {
+            return state != State::STOPPED;
+        }
+
+        void capture_name() {
+            char buf[128];
+            int rc = pthread_getname_np(pthread_id, buf, sizeof(buf));
+            if (rc == 0)
+                name = std::string(buf);
+        }
+};
+
+class Marker {
+    public:
+    enum Type {
+        MARKER_GVL_THREAD_STARTED,
+        MARKER_GVL_THREAD_READY,
+        MARKER_GVL_THREAD_RESUMED,
+        MARKER_GVL_THREAD_SUSPENDED,
+        MARKER_GVL_THREAD_EXITED,
+
+        MARKER_GC_START,
+        MARKER_GC_END_MARK,
+        MARKER_GC_END_SWEEP,
+        MARKER_GC_ENTER,
+        MARKER_GC_EXIT,
+
+        MARKER_MAX,
+    };
+    Type type;
+    TimeStamp timestamp;
+    native_thread_id_t thread_id;
+};
+
+class MarkerTable {
+    public:
+        std::vector<Marker> list;
+        std::mutex mutex;
+
+        void record(Marker::Type type) {
+            const std::lock_guard<std::mutex> lock(mutex);
+
+            list.push_back({ type, TimeStamp::Now(), Thread::get_native_thread_id() });
+        }
+};
+
+extern "C" int ruby_thread_has_gvl_p(void);
+
+class ThreadTable {
+    public:
+        std::vector<Thread> list;
+        std::mutex mutex;
+
+        void started() {
+            //const std::lock_guard<std::mutex> lock(mutex);
+
+            //list.push_back(Thread{pthread_self(), Thread::State::SUSPENDED});
+            set_state(Thread::State::STARTED);
+        }
+
+        void set_state(Thread::State new_state) {
+            const std::lock_guard<std::mutex> lock(mutex);
+
+            pthread_t current_thread = pthread_self();
+            //cerr << "set state=" << new_state << " thread=" << gettid() << endl;
+
+            for (auto &thread : list) {
+                if (pthread_equal(current_thread, thread.pthread_id)) {
+                    thread.set_state(new_state);
+
+                    if (new_state == Thread::State::SUSPENDED) {
+                        thread.stack_on_suspend.sample();
+                        //cerr << gettid() << " suspended! Stack size:" << thread.stack_on_suspend.size() << endl;
+                    }
+                    return;
+                }
+            }
+
+            pid_t native_tid = Thread::get_native_thread_id();
+            list.emplace_back(new_state);
+        }
+};
+
+enum Category{
+	CATEGORY_NORMAL,
+	CATEGORY_IDLE
+};
+
+class TimeCollector : public BaseCollector {
+    std::vector<int> samples;
+    std::vector<TimeStamp> timestamps;
+    std::vector<native_thread_id_t> sample_threads;
+    std::vector<Category> sample_categories;
+
+    MarkerTable markers;
+    ThreadTable threads;
+
+    pthread_t sample_thread;
+
+    atomic_bool running;
+    SamplerSemaphore thread_stopped;
+
+    static inline LiveSample *live_sample;
+
+    TimeStamp started_at;
+    TimeStamp interval;
+
+    public:
+    TimeCollector(TimeStamp interval) : interval(interval) {
+    }
+
+    private:
+
+    void record_sample(const RawSample &sample, TimeStamp time, const Thread &thread, Category category) {
+        if (!sample.empty()) {
+            int stack_index = frame_list.stack_index(sample);
+            samples.push_back(stack_index);
+            timestamps.push_back(time);
+            sample_threads.push_back(thread.native_tid);
+            sample_categories.push_back(category);
+        }
+    }
+
+    static void signal_handler(int sig, siginfo_t* sinfo, void* ucontext) {
+        assert(live_sample);
+        live_sample->sample_current_thread();
+    }
+
+    VALUE get_markers() {
+        VALUE list = rb_ary_new();
+
+        for (auto& marker: this->markers.list) {
+            VALUE record[3] = {0};
+            record[0] = ULL2NUM(marker.thread_id);
+            record[1] = INT2NUM(marker.type);
+            record[2] = ULL2NUM(marker.timestamp.nanoseconds());
+            rb_ary_push(list, rb_ary_new_from_values(3, record));
+        }
+
+        return list;
+    }
+
+    void sample_thread_run() {
+        LiveSample sample;
+        live_sample = &sample;
+
+        TimeStamp next_sample_schedule = TimeStamp::Now();
+        while (running) {
+            TimeStamp sample_start = TimeStamp::Now();
+
+            threads.mutex.lock();
+            for (auto thread : threads.list) {
+                //if (thread.state == Thread::State::RUNNING) {
+                if (thread.state == Thread::State::RUNNING || (thread.state == Thread::State::SUSPENDED && thread.stack_on_suspend.size() == 0)) {
+                    if (pthread_kill(thread.pthread_id, SIGPROF)) {
+                        rb_bug("pthread_kill failed");
+                    }
+                    sample.wait();
+
+                    if (sample.sample.gc) {
+                        // fprintf(stderr, "skipping GC sample\n");
+                    } else {
+                        record_sample(sample.sample, sample_start, thread, CATEGORY_NORMAL);
+                    }
+                } else if (thread.state == Thread::State::SUSPENDED) {
+                    record_sample(thread.stack_on_suspend, sample_start, thread, CATEGORY_IDLE);
+                } else {
+                }
+            }
+            threads.mutex.unlock();
+
+            TimeStamp sample_complete = TimeStamp::Now();
+
+            next_sample_schedule += interval;
+
+            if (next_sample_schedule < sample_complete) {
+                //fprintf(stderr, "fell behind by %ius\n", (sample_complete - next_sample_schedule).microseconds());
+                next_sample_schedule = sample_complete + interval;
+            }
+
+            TimeStamp sleep_time = next_sample_schedule - sample_complete;
+            TimeStamp::Sleep(sleep_time);
+        }
+
+        live_sample = NULL;
+
+        thread_stopped.post();
+    }
+
+    static void *sample_thread_entry(void *arg) {
+        TimeCollector *collector = static_cast<TimeCollector *>(arg);
+        collector->sample_thread_run();
+        return NULL;
+    }
+
+    static void internal_gc_event_cb(VALUE tpval, void *data) {
+        TimeCollector *collector = static_cast<TimeCollector *>(data);
+        rb_trace_arg_t *tparg = rb_tracearg_from_tracepoint(tpval);
+        int event = rb_tracearg_event_flag(tparg);
+
+        switch (event) {
+            case RUBY_INTERNAL_EVENT_GC_START:
+                collector->markers.record(Marker::Type::MARKER_GC_START);
+                break;
+            case RUBY_INTERNAL_EVENT_GC_END_MARK:
+                collector->markers.record(Marker::Type::MARKER_GC_END_MARK);
+                break;
+            case RUBY_INTERNAL_EVENT_GC_END_SWEEP:
+                collector->markers.record(Marker::Type::MARKER_GC_END_SWEEP);
+                break;
+            case RUBY_INTERNAL_EVENT_GC_ENTER:
+                collector->markers.record(Marker::Type::MARKER_GC_ENTER);
+                break;
+            case RUBY_INTERNAL_EVENT_GC_EXIT:
+                collector->markers.record(Marker::Type::MARKER_GC_EXIT);
+                break;
+        }
+    }
+
+    static void internal_thread_event_cb(rb_event_flag_t event, const rb_internal_thread_event_data_t *event_data, void *data) {
+        TimeCollector *collector = static_cast<TimeCollector *>(data);
+        //cerr << "internal thread event" << event << " at " << TimeStamp::Now() << endl;
+
+        switch (event) {
+            case RUBY_INTERNAL_THREAD_EVENT_STARTED:
+                collector->markers.record(Marker::Type::MARKER_GVL_THREAD_STARTED);
+                collector->threads.started();
+                break;
+            case RUBY_INTERNAL_THREAD_EVENT_READY:
+                collector->markers.record(Marker::Type::MARKER_GVL_THREAD_READY);
+                break;
+            case RUBY_INTERNAL_THREAD_EVENT_RESUMED:
+                collector->markers.record(Marker::Type::MARKER_GVL_THREAD_RESUMED);
+                collector->threads.set_state(Thread::State::RUNNING);
+                break;
+            case RUBY_INTERNAL_THREAD_EVENT_SUSPENDED:
+                collector->markers.record(Marker::Type::MARKER_GVL_THREAD_SUSPENDED);
+                collector->threads.set_state(Thread::State::SUSPENDED);
+                break;
+            case RUBY_INTERNAL_THREAD_EVENT_EXITED:
+                collector->markers.record(Marker::Type::MARKER_GVL_THREAD_EXITED);
+                collector->threads.set_state(Thread::State::STOPPED);
+                break;
+
+        }
+    }
+
+    rb_internal_thread_event_hook_t *thread_hook;
+    VALUE gc_hook;
+
+    bool start() {
+        if (!BaseCollector::start()) {
+            return false;
+        }
+
+	started_at = TimeStamp::Now();
+
+        struct sigaction sa;
+        sa.sa_sigaction = signal_handler;
+        sa.sa_flags = SA_RESTART | SA_SIGINFO;
+        sigemptyset(&sa.sa_mask);
+        sigaction(SIGPROF, &sa, NULL);
+
+        running = true;
+
+        int ret = pthread_create(&sample_thread, NULL, &sample_thread_entry, this);
+        if (ret != 0) {
+            perror("pthread_create");
+            rb_bug("pthread_create");
+        }
+
+        // Set the state of the current Ruby thread to RUNNING.
+        // We want to have at least one thread in our thread list because it's
+        // possible that the profile might be such that we don't get any
+        // thread switch events and we need at least one
+        this->threads.set_state(Thread::State::RUNNING);
+
+        thread_hook = rb_internal_thread_add_event_hook(internal_thread_event_cb, RUBY_INTERNAL_THREAD_EVENT_MASK, this);
+        gc_hook = rb_tracepoint_new(0, RUBY_GC_PHASE_EVENTS, internal_gc_event_cb, (void *)this);
+        rb_tracepoint_enable(gc_hook);
+
+        return true;
+    }
+
+    VALUE stop() {
+        BaseCollector::stop();
+
+        running = false;
+        thread_stopped.wait();
+
+        struct sigaction sa;
+        sa.sa_handler = SIG_IGN;
+        sa.sa_flags = SA_RESTART;
+        sigemptyset(&sa.sa_mask);
+        sigaction(SIGPROF, &sa, NULL);
+
+        rb_internal_thread_remove_event_hook(thread_hook);
+        rb_tracepoint_disable(gc_hook);
+
+        // capture thread names
+        for (auto& thread: this->threads.list) {
+            if (thread.running()) {
+                thread.capture_name();
+            }
+        }
+
+        frame_list.finalize();
+
+        VALUE result = build_collector_result();
+
+        reset();
+
+        return result;
+    }
+
+    VALUE build_collector_result() {
+        VALUE result = rb_obj_alloc(rb_cVernierResult);
+
+        VALUE meta = rb_hash_new();
+        rb_ivar_set(result, rb_intern("@meta"), meta);
+        rb_hash_aset(meta, sym("started_at"), ULL2NUM(started_at.nanoseconds()));
+
+        VALUE samples = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@samples"), samples);
+        VALUE weights = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@weights"), weights);
+        for (auto& stack_index: this->samples) {
+            rb_ary_push(samples, INT2NUM(stack_index));
+            rb_ary_push(weights, INT2NUM(1));
+        }
+
+        VALUE timestamps = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@timestamps"), timestamps);
+
+        for (auto& timestamp: this->timestamps) {
+            rb_ary_push(timestamps, ULL2NUM(timestamp.nanoseconds()));
+        }
+
+        VALUE sample_threads = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@sample_threads"), sample_threads);
+        for (auto& thread: this->sample_threads) {
+            rb_ary_push(sample_threads, ULL2NUM(thread));
+        }
+
+        VALUE sample_categories = rb_ary_new();
+        rb_ivar_set(result, rb_intern("@sample_categories"), sample_categories);
+        for (auto& cat: this->sample_categories) {
+            rb_ary_push(sample_categories, INT2NUM(cat));
+        }
+
+        VALUE threads = rb_hash_new();
+        rb_ivar_set(result, rb_intern("@threads"), threads);
+
+        for (const auto& thread: this->threads.list) {
+            VALUE hash = rb_hash_new();
+            rb_hash_aset(threads, ULL2NUM(thread.native_tid), hash);
+            rb_hash_aset(hash, sym("tid"), ULL2NUM(thread.native_tid));
+            rb_hash_aset(hash, sym("started_at"), ULL2NUM(thread.started_at.nanoseconds()));
+            if (!thread.stopped_at.zero()) {
+                rb_hash_aset(hash, sym("stopped_at"), ULL2NUM(thread.stopped_at.nanoseconds()));
+            }
+            rb_hash_aset(hash, sym("name"), rb_str_new(thread.name.data(), thread.name.length()));
+
+        }
+
+        frame_list.write_result(result);
+
+        return result;
+    }
+
+    void mark() {
+        frame_list.mark_frames();
+        rb_gc_mark(gc_hook);
+
+        //for (int i = 0; i < queued_length; i++) {
+        //    rb_gc_mark(queued_frames[i]);
+        //}
+
+        // FIXME: How can we best mark buffered or pending frames?
+    }
+};
 
 static void
-freeobj_i(VALUE tpval, void *data) {
-    retained_collector *collector = static_cast<retained_collector *>(data);
-    TraceArg tp(tpval);
-    collector->freed_objects++;
+collector_mark(void *data) {
+    BaseCollector *collector = static_cast<BaseCollector *>(data);
+    collector->mark();
+}
 
-    collector->object_frames.erase(tp.obj);
+static void
+collector_free(void *data) {
+    BaseCollector *collector = static_cast<BaseCollector *>(data);
+    delete collector;
 }
 
+static const rb_data_type_t rb_collector_type = {
+    .wrap_struct_name = "vernier/collector",
+    .function = {
+        //.dmemsize = rb_collector_memsize,
+        .dmark = collector_mark,
+        .dfree = collector_free,
+    },
+};
 
-static VALUE
-trace_retained_start(VALUE self) {
-    retained_collector *collector = &_collector;
+static BaseCollector *get_collector(VALUE obj) {
+    BaseCollector *collector;
+    TypedData_Get_Struct(obj, BaseCollector, &rb_collector_type, collector);
+    return collector;
+}
 
-    tp_newobj = rb_tracepoint_new(0, RUBY_INTERNAL_EVENT_NEWOBJ, newobj_i, collector);
-    tp_freeobj = rb_tracepoint_new(0, RUBY_INTERNAL_EVENT_FREEOBJ, freeobj_i, collector);
+static VALUE
+collector_start(VALUE self) {
+    auto *collector = get_collector(self);
 
-    rb_tracepoint_enable(tp_newobj);
-    rb_tracepoint_enable(tp_freeobj);
+    if (!collector->start()) {
+        rb_raise(rb_eRuntimeError, "already running");
+    }
 
     return Qtrue;
 }
 
-#define sym(name) ID2SYM(rb_intern_const(name))
-
-// HACK: This isn't public, but the objspace ext uses it
-extern "C" size_t rb_obj_memsize_of(VALUE);
+static VALUE
+collector_stop(VALUE self) {
+    auto *collector = get_collector(self);
 
-static const char *
-ruby_object_type_name(VALUE obj) {
-    enum ruby_value_type type = rb_type(obj);
-
-#define TYPE_CASE(x) case (x): return (#x)
-
-    // Many of these are impossible, but it's easier to just include them
-    switch (type) {
-        TYPE_CASE(T_OBJECT);
-        TYPE_CASE(T_CLASS);
-        TYPE_CASE(T_MODULE);
-        TYPE_CASE(T_FLOAT);
-        TYPE_CASE(T_STRING);
-        TYPE_CASE(T_REGEXP);
-        TYPE_CASE(T_ARRAY);
-        TYPE_CASE(T_HASH);
-        TYPE_CASE(T_STRUCT);
-        TYPE_CASE(T_BIGNUM);
-        TYPE_CASE(T_FILE);
-        TYPE_CASE(T_DATA);
-        TYPE_CASE(T_MATCH);
-        TYPE_CASE(T_COMPLEX);
-        TYPE_CASE(T_RATIONAL);
-
-        TYPE_CASE(T_NIL);
-        TYPE_CASE(T_TRUE);
-        TYPE_CASE(T_FALSE);
-        TYPE_CASE(T_SYMBOL);
-        TYPE_CASE(T_FIXNUM);
-        TYPE_CASE(T_UNDEF);
-
-        TYPE_CASE(T_IMEMO);
-        TYPE_CASE(T_NODE);
-        TYPE_CASE(T_ICLASS);
-        TYPE_CASE(T_ZOMBIE);
-        TYPE_CASE(T_MOVED);
-
-        default:
-        return "unknown type";
-    }
-#undef TYPE_CASE
+    VALUE result = collector->stop();
+    return result;
 }
 
 static VALUE
-trace_retained_stop(VALUE self) {
-    rb_tracepoint_disable(tp_newobj);
-    rb_tracepoint_disable(tp_freeobj);
+markers(VALUE self) {
+    auto *collector = get_collector(self);
 
-    retained_collector *collector = &_collector;
+    return collector->get_markers();
+}
 
-    std::stringstream ss;
+static VALUE
+collector_sample(VALUE self) {
+    auto *collector = get_collector(self);
 
-    for (auto& it: collector->object_frames) {
-        VALUE obj = it.first;
-        const Stack &stack = *it.second;
+    collector->sample();
+    return Qtrue;
+}
 
-        for (int i = stack.size() - 1; i >= 0; i--) {
-            const Frame &frame = stack.frame(i);
-            ss << frame;
-            if (i > 0) ss << ";";
+static VALUE collector_new(VALUE self, VALUE mode, VALUE options) {
+    BaseCollector *collector;
+    if (mode == sym("retained")) {
+        collector = new RetainedCollector();
+    } else if (mode == sym("custom")) {
+        collector = new CustomCollector();
+    } else if (mode == sym("wall")) {
+        VALUE intervalv = rb_hash_aref(options, sym("interval"));
+        TimeStamp interval;
+        if (NIL_P(intervalv)) {
+            interval = TimeStamp::from_microseconds(500);
+        } else {
+            interval = TimeStamp::from_microseconds(NUM2UINT(intervalv));
         }
-        ss << ";" << ruby_object_type_name(obj);
-        ss << " " << rb_obj_memsize_of(obj) << endl;
+        collector = new TimeCollector(interval);
+    } else {
+        rb_raise(rb_eArgError, "invalid mode");
     }
-
-    std::string s = ss.str();
-    VALUE str = rb_str_new(s.c_str(), s.size());
-
-    return str;
+    VALUE obj = TypedData_Wrap_Struct(self, &rb_collector_type, collector);
+    rb_funcall(obj, rb_intern("initialize"), 1, mode);
+    return obj;
 }
 
 static void
-retained_collector_mark(void *data) {
-    retained_collector *collector = static_cast<retained_collector *>(data);
+Init_consts() {
+#define MARKER_CONST(name) \
+    rb_define_const(rb_mVernierMarkerType, #name, INT2NUM(Marker::Type::MARKER_##name))
 
-    // We don't mark the objects, but we MUST mark the frames, otherwise they
-    // can be garbage collected.
-    // This may lead to method entries being unnecessarily retained.
-    for (VALUE frame: collector->unique_frames) {
-        rb_gc_mark(frame);
-    }
+    MARKER_CONST(GVL_THREAD_STARTED);
+    MARKER_CONST(GVL_THREAD_READY);
+    MARKER_CONST(GVL_THREAD_RESUMED);
+    MARKER_CONST(GVL_THREAD_SUSPENDED);
+    MARKER_CONST(GVL_THREAD_EXITED);
+
+    MARKER_CONST(GC_START);
+    MARKER_CONST(GC_END_MARK);
+    MARKER_CONST(GC_END_SWEEP);
+    MARKER_CONST(GC_ENTER);
+    MARKER_CONST(GC_EXIT);
+
+#undef MARKER_CONST
 }
 
 extern "C" void
 Init_vernier(void)
 {
   rb_mVernier = rb_define_module("Vernier");
+  rb_cVernierResult = rb_define_class_under(rb_mVernier, "Result", rb_cObject);
+  VALUE rb_mVernierMarker = rb_define_module_under(rb_mVernier, "Marker");
+  rb_mVernierMarkerType = rb_define_module_under(rb_mVernierMarker, "Type");
+
+  rb_cVernierCollector = rb_define_class_under(rb_mVernier, "Collector", rb_cObject);
+  rb_undef_alloc_func(rb_cVernierCollector);
+  rb_define_singleton_method(rb_cVernierCollector, "_new", collector_new, 2);
+  rb_define_method(rb_cVernierCollector, "start", collector_start, 0);
+  rb_define_method(rb_cVernierCollector, "sample", collector_sample, 0);
+  rb_define_private_method(rb_cVernierCollector, "finish",  collector_stop, 0);
+  rb_define_private_method(rb_cVernierCollector, "markers",  markers, 0);
 
-  rb_define_module_function(rb_mVernier, "trace_retained_start", trace_retained_start, 0);
-  rb_define_module_function(rb_mVernier, "trace_retained_stop", trace_retained_stop, 0);
+  Init_consts();
 
-  static VALUE gc_hook = Data_Wrap_Struct(rb_cObject, retained_collector_mark, NULL, &_collector);
-  rb_global_variable(&gc_hook);
+  //static VALUE gc_hook = Data_Wrap_Struct(rb_cObject, collector_mark, NULL, &_collector);
+  //rb_global_variable(&gc_hook);
 }