move TaskStatsData to OSS dynolog (#469)

Summary:

TaskStatsData.h contains most of the data models used by ThreadMonitor. This is an essential step In order to finally create an OSS ThreadMonitor.

Differential Revision: D86358725

Author: yaoz08

dynolog/src/threadmon/TaskStatsData.h

@@ -0,0 +1,259 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <stdint.h>
+#include <sys/types.h>
+#include <algorithm>
+#include <chrono>
+#include <string>
+
+namespace dynolog {
+
+using timestamp = std::chrono::time_point<std::chrono::steady_clock>;
+
+struct TaskStat {
+  time_t t = 0;
+  timestamp ts;
+  uint32_t seq = 0;
+
+  pid_t tid = 0;
+  pid_t ppid = 0;
+  std::string comm;
+
+  std::string tcomm;
+
+  uint64_t cpuCount = 0; /* number of schedules */
+  uint64_t cpuDelayTotalNs = 0; /* ns waiting to be scheduled */
+  uint64_t cpuRunRealTotalNs = 0; /* total real cpu running time */
+  uint64_t acMinflt = 0; /* Minor Page Fault Count - copy on write */
+  uint64_t acMajflt = 0; /* Major Page Fault Count - virtual memory */
+  /* Delay waiting for page fault I/O (swap in only) */
+  uint64_t swapinCount = 0;
+  uint64_t swapinDelayTotal = 0;
+  uint64_t voluntaryCSW = 0; /* Voluntary Context Switches */
+  uint64_t involuntaryCSW = 0; /* Involuntary Context Switches */
+  uint64_t etime = 0; /* elapsed time in us */
+  uint64_t utime = 0; /* User cpu time in us */
+  uint64_t stime = 0; /* System cpu time in us */
+  int64_t memdelayUs = 0; /* type waiting for memory page */
+  uint64_t io_read_bytes = 0; /* bytes of read i/o */
+  uint64_t io_write_bytes = 0; /* bytes of write i/o */
+  /* High watermark of RSS usage in duration of a task, in KBytes. */
+  uint64_t hiwaterRssKb = 0; /* High-watermark of RSS usage in KBytes */
+};
+
+// calculate difference assuming possible 32-bit counter overflow
+static inline uint64_t deltaWrap32(uint64_t cur, uint64_t prev) {
+  return cur < prev ? (UINT32_MAX - prev) + cur : cur - prev;
+}
+
+struct TaskStatsDelta {
+  uint64_t cpuCount = 0;
+  uint64_t cpuDelayUs = 0; /* time waiting to be scheduled */
+  uint64_t cpuRunRealUs = 0; /* total real cpu running time */
+  uint64_t acMinflt = 0;
+  uint64_t acMajflt = 0;
+  uint64_t swapinCount = 0;
+  uint64_t swapinDelayUs = 0;
+  uint64_t voluntaryCSW = 0; /* Voluntary Context Switches */
+  uint64_t involuntaryCSW = 0; /* Involuntary Context Switches */
+  uint64_t utimeUs = 0; /* user cpu time in us */
+  uint64_t stimeUs = 0; /* time spent in System */
+  uint64_t etimeUs = 0; /* time elapsed */
+  int64_t memdelayUs = 0; /* time waiting for memory page */
+  uint64_t io_read_bytes = 0; /* bytes of read i/o */
+  uint64_t io_write_bytes = 0; /* bytes of write i/o */
+  uint64_t hiwaterRssKb = 0;
+
+  explicit TaskStatsDelta(const TaskStat& prev, const TaskStat& cur) {
+    // It's quite annoying, but even though taskstats return 64-bit counters,
+    // kernel actually keeps track of some of them in 32-bit counters, some in
+    // 64-bit counters. To add to that confusion, taskstats has additional logic
+    // to not roll-over counters on overflow and truncate it to zero (but that
+    // will be relevant only to per-process stats, which can accumulate large
+    // enough value across many threads, not per-thread stats, because for
+    // single thread signed 64-bit overflow can happen only after 292 years). We
+    // carefully take all that into account here when calculating deltas.
+
+    // cpuCount is unsigned 64-bit counter (good for 584 years)
+    cpuCount = cur.cpuCount - prev.cpuCount;
+
+    // cpuDelay is signed 64-bit counter (good for 292 years)
+    cpuDelayUs = (cur.cpuDelayTotalNs - prev.cpuDelayTotalNs) / 1000;
+
+    // cpuRunReal is signed 64-bit counter (good for 292 years)
+    cpuRunRealUs = (cur.cpuRunRealTotalNs - prev.cpuRunRealTotalNs) / 1000;
+
+    // swapinCount is 32-bit counter
+    swapinCount = deltaWrap32(cur.swapinCount, prev.swapinCount);
+
+    // swapinDelay is unsigned 64-bit counter (good for 584 years)
+    swapinDelayUs = (cur.swapinDelayTotal - prev.swapinDelayTotal) / 1000;
+
+    // faults and context switches are unsigned 64-bit counters
+    acMinflt = cur.acMinflt - prev.acMinflt;
+    acMajflt = cur.acMajflt - prev.acMajflt;
+    voluntaryCSW = cur.voluntaryCSW - prev.voluntaryCSW;
+    involuntaryCSW = cur.involuntaryCSW - prev.involuntaryCSW;
+
+    // etime is 64-bit counter always, we are good for at least 584 years
+    etimeUs = cur.etime - prev.etime;
+
+    // utime/stime are unsigned 64-bit counters tracked as nanoseconds.
+    utimeUs = cur.utime - prev.utime;
+    stimeUs = cur.stime - prev.stime;
+
+    memdelayUs = prev.memdelayUs >= 0 && cur.memdelayUs >= prev.memdelayUs
+        ? cur.memdelayUs - prev.memdelayUs
+        : 0;
+
+    // io reads and writes are unsigned 64-bit counters
+    io_read_bytes = cur.io_read_bytes - prev.io_read_bytes;
+    io_write_bytes = cur.io_write_bytes - prev.io_write_bytes;
+
+    // at least is all straightforward with memory high watermark
+    hiwaterRssKb = cur.hiwaterRssKb > prev.hiwaterRssKb ? cur.hiwaterRssKb
+                                                        : prev.hiwaterRssKb;
+  }
+};
+
+// aggregated thread stats
+struct ThreadSummaryStats {
+  uint32_t activeThreads = 0; /* number of live threads */
+  uint32_t exitedThreads = 0; /* number of exited threads */
+  uint64_t schedThreads = 0; /* number of threads scheduled to be run on CPU */
+
+  uint64_t utimeUs = 0; /* time spent in User */
+  uint64_t stimeUs = 0; /* time spent in System */
+  uint64_t etimeUs = 0; /* time elapsed */
+  uint64_t utimeMaxUs = 0; /* max utimeUs of all threads */
+  uint64_t stimeMaxUs = 0; /* max stimeUs of all threads */
+  uint64_t cpuTimeMaxUs = 0; /* max total cpu time of all threads */
+
+  uint64_t hiwaterRssKb = 0; /* maximum RSS in KB */
+
+  uint64_t schedDelayUs = 0; /* time waiting to be scheduled */
+  uint64_t schedCount = 0; /* number of times thread was scheduled */
+  uint64_t voluntaryCSW = 0; /* voluntary context switches */
+  uint64_t involuntaryCSW = 0; /* involuntary context switches */
+  uint64_t acMinflt = 0; /* minor faults */
+  uint64_t acMajflt = 0; /* major faults */
+  uint64_t swapinCount = 0; /* number of memory swap ins */
+  uint64_t swapinDelayUs = 0; /* memory swap in elapsed time */
+  int64_t memdelayUs = 0; /* amount of time thread waited to get memory */
+
+  uint64_t io_read_bytes = 0; /* bytes of read i/o */
+  uint64_t io_write_bytes = 0; /* bytes of write i/o */
+
+  void aggregate(const TaskStatsDelta& d, bool exited) {
+    activeThreads += exited ? 0 : 1;
+    exitedThreads += exited ? 1 : 0;
+    schedThreads += d.cpuCount > 0 ? 1 : 0;
+
+    utimeUs += d.utimeUs;
+    stimeUs += d.stimeUs;
+    etimeUs += d.etimeUs;
+    utimeMaxUs = std::max(utimeMaxUs, d.utimeUs);
+    stimeMaxUs = std::max(stimeMaxUs, d.stimeUs);
+    cpuTimeMaxUs = std::max(cpuTimeMaxUs, d.utimeUs + d.stimeUs);
+
+    hiwaterRssKb =
+        hiwaterRssKb > d.hiwaterRssKb ? hiwaterRssKb : d.hiwaterRssKb;
+
+    schedDelayUs += d.cpuDelayUs;
+    schedCount += d.cpuCount;
+    voluntaryCSW += d.voluntaryCSW;
+    involuntaryCSW += d.involuntaryCSW;
+    acMinflt += d.acMinflt;
+    acMajflt += d.acMajflt;
+    swapinCount += d.swapinCount;
+    swapinDelayUs += d.swapinDelayUs;
+    memdelayUs += d.memdelayUs;
+
+    io_read_bytes += d.io_read_bytes;
+    io_write_bytes += d.io_write_bytes;
+  }
+
+  void aggregate(const ThreadSummaryStats& s) {
+    activeThreads += s.activeThreads;
+    exitedThreads += s.exitedThreads;
+    schedThreads += s.schedThreads;
+
+    utimeUs += s.utimeUs;
+    stimeUs += s.stimeUs;
+    etimeUs += s.etimeUs;
+    utimeMaxUs = std::max(utimeMaxUs, s.utimeMaxUs);
+    stimeMaxUs = std::max(stimeMaxUs, s.stimeMaxUs);
+    cpuTimeMaxUs = std::max(cpuTimeMaxUs, s.cpuTimeMaxUs);
+
+    hiwaterRssKb =
+        hiwaterRssKb > s.hiwaterRssKb ? hiwaterRssKb : s.hiwaterRssKb;
+
+    schedDelayUs += s.schedDelayUs;
+    schedCount += s.schedCount;
+    voluntaryCSW += s.voluntaryCSW;
+    involuntaryCSW += s.involuntaryCSW;
+    acMinflt += s.acMinflt;
+    acMajflt += s.acMajflt;
+    swapinCount += s.swapinCount;
+    swapinDelayUs += s.swapinDelayUs;
+    memdelayUs += s.memdelayUs;
+
+    io_read_bytes += s.io_read_bytes;
+    io_write_bytes += s.io_write_bytes;
+  }
+};
+
+// counters maintained by ThreadMonitor
+// calculated from aggregating thread-level data
+struct ThreadAggCounters {
+  uint64_t sumEtimeUs = 0;
+  uint64_t countEtimeUs = 0;
+  uint8_t maxThreadCpuUtil = 0;
+  unsigned int swapinCount = 0;
+  unsigned int swapinDelayUs = 0;
+  unsigned int taskCount = 0;
+  unsigned int taskActiveCount = 0;
+  uint64_t majorFaultCount = 0;
+  uint64_t minorFaultCount = 0;
+  uint64_t voluntaryCSW = 0;
+  uint64_t involuntaryCSW = 0;
+  uint64_t cpuDelayUs = 0;
+  uint64_t cpuRunRealUs = 0;
+
+  void updateThreadAggCounters(const TaskStatsDelta& delta) {
+    sumEtimeUs += delta.etimeUs;
+    countEtimeUs++;
+
+    float cpuUtil = 0.0;
+    if (delta.etimeUs > 0) {
+      auto util =
+          100.0 * (delta.utimeUs + delta.stimeUs) / (delta.etimeUs * 1.0);
+      // utime and stime may be rounded to the closest ms while etime remains
+      // accurate, this could result in a cpu util above 100%, for now no clear
+      // way to avoid this in the bfp probe so clamp down the value to 100%
+      cpuUtil = (util > 100.0) ? 100.0 : util;
+    }
+    maxThreadCpuUtil = std::max(maxThreadCpuUtil, (uint8_t)cpuUtil);
+    swapinCount += delta.swapinCount;
+    swapinDelayUs += delta.swapinCount;
+    majorFaultCount += delta.acMajflt;
+    minorFaultCount += delta.acMinflt;
+    voluntaryCSW += delta.voluntaryCSW;
+    involuntaryCSW += delta.involuntaryCSW;
+    cpuDelayUs += delta.cpuDelayUs;
+    cpuRunRealUs += delta.cpuRunRealUs;
+    taskCount++;
+    if (delta.cpuCount) {
+      taskActiveCount++;
+    }
+  }
+};
+
+} // namespace dynolog