5 files changed, 325 insertions, 0 deletions
diff --git a/src/crimson/thread/Condition.h b/src/crimson/thread/Condition.h
new file mode 100644
index 00000000..2a5c643d
--- /dev/null
+++ b/src/crimson/thread/Condition.h
@@ -0,0 +1,36 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#pragma once
+
+#include <seastar/core/reactor.hh>
+#include <sys/eventfd.h>
+
+namespace ceph::thread {
+
+/// a synchronization primitive can be used to block a seastar thread, until
+/// another thread notifies it.
+class Condition {
+  seastar::file_desc file_desc;
+  int fd;
+  seastar::pollable_fd_state fd_state;
+  eventfd_t event = 0;
+public:
+  Condition()
+    : file_desc{seastar::file_desc::eventfd(0, 0)},
+      fd(file_desc.get()),
+      fd_state{std::move(file_desc)}
+  {}
+  seastar::future<> wait() {
+    return seastar::engine().read_some(fd_state, &event, sizeof(event))
+      .then([](size_t) {
+	  return seastar::now();
+	});
+  }
+  void notify() {
+    eventfd_t result = 1;
+    ::eventfd_write(fd, result);
+  }
+};
+
+} // namespace ceph::thread
diff --git a/src/crimson/thread/ThreadPool.cc b/src/crimson/thread/ThreadPool.cc
new file mode 100644
index 00000000..9df849b5
--- /dev/null
+++ b/src/crimson/thread/ThreadPool.cc
@@ -0,0 +1,76 @@
+#include "ThreadPool.h"
+
+#include <pthread.h>
+#include "crimson/net/Config.h"
+#include "include/intarith.h"
+
+#include "include/ceph_assert.h"
+
+namespace ceph::thread {
+
+ThreadPool::ThreadPool(size_t n_threads,
+                       size_t queue_sz,
+                       unsigned cpu_id)
+  : queue_size{round_up_to(queue_sz, seastar::smp::count)},
+    pending{queue_size}
+{
+  for (size_t i = 0; i < n_threads; i++) {
+    threads.emplace_back([this, cpu_id] {
+      pin(cpu_id);
+      loop();
+    });
+  }
+}
+
+ThreadPool::~ThreadPool()
+{
+  for (auto& thread : threads) {
+    thread.join();
+  }
+}
+
+void ThreadPool::pin(unsigned cpu_id)
+{
+  cpu_set_t cs;
+  CPU_ZERO(&cs);
+  CPU_SET(cpu_id, &cs);
+  [[maybe_unused]] auto r = pthread_setaffinity_np(pthread_self(),
+                                                   sizeof(cs), &cs);
+  ceph_assert(r == 0);
+}
+
+void ThreadPool::loop()
+{
+  for (;;) {
+    WorkItem* work_item = nullptr;
+    {
+      std::unique_lock lock{mutex};
+      cond.wait_for(lock,
+                    ceph::net::conf.threadpool_empty_queue_max_wait,
+                    [this, &work_item] {
+        return pending.pop(work_item) || is_stopping();
+      });
+    }
+    if (work_item) {
+      work_item->process();
+    } else if (is_stopping()) {
+      break;
+    }
+  }
+}
+
+seastar::future<> ThreadPool::start()
+{
+  auto slots_per_shard = queue_size / seastar::smp::count;
+  return submit_queue.start(slots_per_shard);
+}
+
+seastar::future<> ThreadPool::stop()
+{
+  return submit_queue.stop().then([this] {
+    stopping = true;
+    cond.notify_all();
+  });
+}
+
+} // namespace ceph::thread
diff --git a/src/crimson/thread/ThreadPool.h b/src/crimson/thread/ThreadPool.h
new file mode 100644
index 00000000..cfd72d2a
--- /dev/null
+++ b/src/crimson/thread/ThreadPool.h
@@ -0,0 +1,118 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+#pragma once
+
+#include <atomic>
+#include <condition_variable>
+#include <tuple>
+#include <type_traits>
+#include <boost/lockfree/queue.hpp>
+#include <boost/optional.hpp>
+#include <seastar/core/future.hh>
+#include <seastar/core/gate.hh>
+#include <seastar/core/semaphore.hh>
+#include <seastar/core/sharded.hh>
+
+#include "Condition.h"
+
+namespace ceph::thread {
+
+struct WorkItem {
+  virtual ~WorkItem() {}
+  virtual void process() = 0;
+};
+
+template<typename Func, typename T = std::invoke_result_t<Func>>
+struct Task final : WorkItem {
+  Func func;
+  seastar::future_state<T> state;
+  ceph::thread::Condition on_done;
+public:
+  explicit Task(Func&& f)
+    : func(std::move(f))
+  {}
+  void process() override {
+    try {
+      state.set(func());
+    } catch (...) {
+      state.set_exception(std::current_exception());
+    }
+    on_done.notify();
+  }
+  seastar::future<T> get_future() {
+    return on_done.wait().then([this] {
+      return seastar::make_ready_future<T>(state.get0(std::move(state).get()));
+    });
+  }
+};
+
+struct SubmitQueue {
+  seastar::semaphore free_slots;
+  seastar::gate pending_tasks;
+  explicit SubmitQueue(size_t num_free_slots)
+    : free_slots(num_free_slots)
+  {}
+  seastar::future<> stop() {
+    return pending_tasks.close();
+  }
+};
+
+/// an engine for scheduling non-seastar tasks from seastar fibers
+class ThreadPool {
+  std::atomic<bool> stopping = false;
+  std::mutex mutex;
+  std::condition_variable cond;
+  std::vector<std::thread> threads;
+  seastar::sharded<SubmitQueue> submit_queue;
+  const size_t queue_size;
+  boost::lockfree::queue<WorkItem*> pending;
+
+  void loop();
+  bool is_stopping() const {
+    return stopping.load(std::memory_order_relaxed);
+  }
+  static void pin(unsigned cpu_id);
+  seastar::semaphore& local_free_slots() {
+    return submit_queue.local().free_slots;
+  }
+  ThreadPool(const ThreadPool&) = delete;
+  ThreadPool& operator=(const ThreadPool&) = delete;
+public:
+  /**
+   * @param queue_sz the depth of pending queue. before a task is scheduled,
+   *                 it waits in this queue. we will round this number to
+   *                 multiple of the number of cores.
+   * @param n_threads the number of threads in this thread pool.
+   * @param cpu the CPU core to which this thread pool is assigned
+   * @note each @c Task has its own ceph::thread::Condition, which possesses
+   * possesses an fd, so we should keep the size of queue under a reasonable
+   * limit.
+   */
+  ThreadPool(size_t n_threads, size_t queue_sz, unsigned cpu);
+  ~ThreadPool();
+  seastar::future<> start();
+  seastar::future<> stop();
+  template<typename Func, typename...Args>
+  auto submit(Func&& func, Args&&... args) {
+    auto packaged = [func=std::move(func),
+                     args=std::forward_as_tuple(args...)] {
+      return std::apply(std::move(func), std::move(args));
+    };
+    return seastar::with_gate(submit_queue.local().pending_tasks,
+      [packaged=std::move(packaged), this] {
+        return local_free_slots().wait()
+          .then([packaged=std::move(packaged), this] {
+            auto task = new Task{std::move(packaged)};
+            auto fut = task->get_future();
+            pending.push(task);
+            cond.notify_one();
+            return fut.finally([task, this] {
+              local_free_slots().signal();
+              delete task;
+            });
+          });
+        });
+  }
+};
+
+} // namespace ceph::thread
diff --git a/src/crimson/thread/Throttle.cc b/src/crimson/thread/Throttle.cc
new file mode 100644
index 00000000..1d67e723
--- /dev/null
+++ b/src/crimson/thread/Throttle.cc
@@ -0,0 +1,59 @@
+#include "Throttle.h"
+
+namespace ceph::thread {
+
+int64_t Throttle::take(int64_t c)
+{
+  if (!max) {
+    return 0;
+  }
+  count += c;
+  return count;
+}
+
+int64_t Throttle::put(int64_t c)
+{
+  if (!max) {
+    return 0;
+  }
+  if (!c) {
+    return count;
+  }
+  on_free_slots.signal();
+  count -= c;
+  return count;
+}
+
+seastar::future<> Throttle::get(size_t c)
+{
+  if (!max) {
+    return seastar::now();
+  }
+  return on_free_slots.wait([this, c] {
+    return !_should_wait(c);
+  }).then([this, c] {
+    count += c;
+    return seastar::now();
+  });
+}
+
+void Throttle::reset_max(size_t m) {
+  if (max == m) {
+    return;
+  }
+
+  if (m > max) {
+    on_free_slots.signal();
+  }
+  max = m;
+}
+
+bool Throttle::_should_wait(size_t c) const {
+  if (!max) {
+    return false;
+  }
+  return ((c <= max && count + c > max) || // normally stay under max
+          (c >= max && count > max));      // except for large c
+}
+
+} // namespace ceph::thread::seastar
diff --git a/src/crimson/thread/Throttle.h b/src/crimson/thread/Throttle.h
new file mode 100644
index 00000000..c2342171
--- /dev/null
+++ b/src/crimson/thread/Throttle.h
@@ -0,0 +1,36 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:nil -*-
+// vim: ts=8 sw=2 smarttab
+
+#pragma once
+
+#include <seastar/core/condition-variable.hh>
+
+#include "common/ThrottleInterface.h"
+
+namespace ceph::thread {
+
+class Throttle final : public ThrottleInterface {
+  size_t max = 0;
+  size_t count = 0;
+  // we cannot change the "count" of seastar::semaphore after it is created,
+  // so use condition_variable instead.
+  seastar::condition_variable on_free_slots;
+public:
+  explicit Throttle(size_t m)
+    : max(m)
+  {}
+  int64_t take(int64_t c = 1) override;
+  int64_t put(int64_t c = 1) override;
+  seastar::future<> get(size_t c);
+  size_t get_current() const {
+    return count;
+  }
+  size_t get_max() const {
+    return max;
+  }
+  void reset_max(size_t m);
+private:
+  bool _should_wait(size_t c) const;
+};
+
+} // namespace ceph::thread