[tools][perf] Adding linux-perf d8 helper

- Start moving profiler scripts to tools/profiling
- Add linux-perf-d8.py wrapper script that runs `perf record` and
  `perf inject`
- Improve waiting for the d8/chrome process and allow for early
  termination if --timeout is provided
- Allow fractional seconds for --timeout
- Delete run-perf.sh and provide equivalent functionality in
  linux-perf-d8.py

Change-Id: Iac1d6cf895aa7159a9bbb387aca7632df27a0ca3
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3585951
Reviewed-by: Leszek Swirski <leszeks@chromium.org>
Commit-Queue: Camillo Bruni <cbruni@chromium.org>
Cr-Commit-Position: refs/heads/main@{#79991}

tools/profiling/README.md

@@ -0,0 +1,8 @@
+# Profiling Tools
+
+This directory contains various helper scripts to assist with profiling
+d8 and chrome.
+
+## linux perf
+ For [linux perf](https://perf.wiki.kernel.org/) support you can find more
+ documentation and instructions at <https://v8.dev/docs/linux-perf>.

tools/profiling/linux-perf-chrome.py

@@ -5,10 +5,8 @@
 
 import optparse
 from pathlib import Path
-from re import A
 import os
 import shlex
-from signal import SIGQUIT
 import subprocess
 import signal
 import tempfile
@@ -16,8 +14,6 @@ import time
 import psutil
 import multiprocessing
 
-from unittest import result
-
 renderer_cmd_file = Path(__file__).parent / 'linux-perf-renderer-cmd.sh'
 assert renderer_cmd_file.is_file()
 renderer_cmd_prefix = f"{renderer_cmd_file} --perf-data-prefix=chrome_renderer"
@@ -46,7 +42,7 @@ parser.add_option(
     help="Also start linux-perf for the browser process. "
     "By default only renderer processes are sampled. "
     "Outputs 'browser_*.perf.data' in the CDW")
-parser.add_option("--timeout", type=int, help="Stop chrome after N seconds")
+parser.add_option("--timeout", type=float, help="Stop chrome after N seconds")
 
 chrome_options = optparse.OptionGroup(
     parser, "Chrome-forwarded Options",
@@ -116,6 +112,17 @@ os.chdir(options.perf_data_dir)
 JS_FLAGS_PERF = ("--perf-prof --no-write-protect-code-memory "
                  "--interpreted-frames-native-stack")
 
+
+def wait_for_process_timeout(process):
+  sleeping_time = 0
+  while (sleeping_time < options.timeout):
+    processHasStopped = process.poll() is not None
+    if processHasStopped:
+      return True
+    time.sleep(0.5)
+  return False
+
+
 with tempfile.TemporaryDirectory(prefix="chrome-") as tmp_dir_path:
   tempdir = Path(tmp_dir_path)
   cmd = [
@@ -151,7 +158,7 @@ with tempfile.TemporaryDirectory(prefix="chrome-") as tmp_dir_path:
     subprocess.run(cmd)
   else:
     process = subprocess.Popen(cmd)
-    time.sleep(options.timeout)
+    if not wait_for_process_timeout(process):
       log(f"QUITING chrome child processes after {options.timeout}s timeout")
     current_process = psutil.Process()
     children = current_process.children(recursive=True)
@@ -202,6 +209,6 @@ for output_file in reversed(results):
       f"{output_file.name:67}{(output_file.stat().st_size*BYTES_TO_MIB):10.2f}MiB"
   )
 
-log("PPROF EXAMPLE")
+log("PPROF")
 path_strings = map(lambda f: str(f.relative_to(old_cwd)), results)
 print(f"pprof -flame { ' '.join(path_strings)}")

tools/profiling/linux-perf-d8.py

@@ -0,0 +1,203 @@
+#!/usr/bin/env python3
+# Copyright 2022 the V8 project authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+
+from datetime import datetime
+from pathlib import Path
+import optparse
+import os
+import psutil
+import shlex
+import signal
+import subprocess
+import time
+
+# ==============================================================================
+
+usage = """Usage: %prog $D8_BIN [OPTION]... -- [D8_OPTION]... [FILE]
+
+This script runs linux-perf with custom V8 logging to get support to resolve
+JS function names.
+
+The perf data is written to OUT_DIR separate by renderer process.
+
+See http://v8.dev//linux-perf for more detailed instructions.
+See $D8_BIN --help for more options
+"""
+parser = optparse.OptionParser(usage=usage)
+parser.add_option(
+    '--perf-data-dir',
+    default=None,
+    metavar="OUT_DIR",
+    help="Output directory for linux perf profile files")
+parser.add_option("--timeout", type=float, help="Stop d8 after N seconds")
+
+d8_options = optparse.OptionGroup(
+    parser, "d8-forwarded Options",
+    "THese options are for a better script experience that are forward directly"
+    "to d8. Any other d8 option can be passed after the '--' arguments"
+    "separator.")
+d8_options.add_option(
+    "--perf-prof-annotate-wasm",
+    help="Load wasm source map and provide annotate wasm code.",
+    action="store_true",
+    default=False)
+d8_options.add_option(
+    "--no-interpreted-frames-native-stack",
+    help="For profiling v8 copies the interpreter entry trampoline for every "
+    "interpreted function. This makes interpreted functions identifiable on the "
+    "native stack at cost of a slight performance and memory overhead.",
+    action="store_true",
+    default=False)
+parser.add_option_group(d8_options)
+
+perf_options = optparse.OptionGroup(
+    parser, "perf-forward options", """
+These options are forward directly to the `perf record` command and can be
+used to manually tweak how profiling works.
+
+See `perf record --help for more` informationn
+""")
+
+perf_options.add_option(
+    "--freq",
+    default="max",
+    help="Sampling frequency, either 'max' or a number in herz. "
+    "Might be reduced depending on the platform.")
+perf_options.add_option("--call-graph", default="fp", help="Defaults tp 'fp'")
+perf_options.add_option("--event", default=None, help="Not set by default.")
+perf_options.add_option(
+    "--raw-samples",
+    default=None,
+    help="Collect raw samples. Not set by default")
+perf_options.add_option(
+    "--count", default=None, help="Event period to sample. Not set by default.")
+perf_options.add_option(
+    "--no-inherit",
+    action="store_true",
+    default=False,
+    help=" Child tasks do not inherit counters.")
+parser.add_option_group(perf_options)
+
+
+# ==============================================================================
+def log(*args):
+  print("")
+  print("=" * 80)
+  print(*args)
+  print("=" * 80)
+
+
+# ==============================================================================
+(options, args) = parser.parse_args()
+
+if len(args) == 0:
+  parser.error("No d8 binary provided")
+
+d8_bin = Path(args.pop(0))
+if not d8_bin.exists():
+  parser.error(f"D8 '{d8_bin}' does not exist")
+
+if options.perf_data_dir is None:
+  options.perf_data_dir = Path.cwd()
+else:
+  options.perf_data_dir = Path(options.perf_data_dir).absolute()
+
+if not options.perf_data_dir.is_dir():
+  parser.error(f"--perf-data-dir={options.perf_data_dir} "
+               "is not an directory or does not exist.")
+
+if options.timeout and options.timeout < 0:
+  parser.error("--timeout should be a positive number")
+
+# ==============================================================================
+old_cwd = Path.cwd()
+os.chdir(options.perf_data_dir)
+
+# ==============================================================================
+cmd = [str(d8_bin), "--perf-prof"]
+
+if not options.no_interpreted_frames_native_stack:
+  cmd += ["--interpreted-frames-native-stack"]
+if options.perf_prof_annotate_wasm:
+  cmd += ["--perf-prof-annotate-wasm"]
+cmd += args
+log("D8 CMD: ", shlex.join(cmd))
+
+datetime_str = datetime.now().strftime("%Y-%m-%d_%H%M%S")
+perf_data_file = Path.cwd() / f"d8_{datetime_str}.perf.data"
+perf_cmd = [
+    "perf", "record", f"--call-graph={options.call_graph}",
+    f"--freq={options.freq}", "--clockid=mono", f"--output={perf_data_file}"
+]
+if options.count:
+  perf_cmd += [f"--count={options.count}"]
+if options.raw_samples:
+  perf_cmd += [f"--raw_samples={options.raw_samples}"]
+if options.event:
+  perf_cmd += [f"--event={options.event}"]
+if options.no_inherit:
+  perf_cmd += [f"--no-inherit"]
+
+cmd = perf_cmd + ["--"] + cmd
+log("LINUX PERF CMD: ", shlex.join(cmd))
+
+
+def wait_for_process_timeout(process):
+  sleeping_time = 0
+  while (sleeping_time < options.timeout):
+    processHasStopped = process.poll() is not None
+    if processHasStopped:
+      return True
+    time.sleep(0.1)
+  return False
+
+
+if options.timeout is None:
+  subprocess.run(cmd)
+else:
+  process = subprocess.Popen(cmd)
+  if not wait_for_process_timeout(process):
+    log(f"QUITING d8 processes after {options.timeout}s timeout")
+  current_process = psutil.Process()
+  children = current_process.children(recursive=True)
+  for child in children:
+    if "d8" in child.name():
+      print(f"  quitting PID={child.pid}")
+      child.send_signal(signal.SIGQUIT)
+  # Wait for linux-perf to write out files
+  time.sleep(1)
+  process.send_signal(signal.SIGQUIT)
+  process.wait()
+
+# ==============================================================================
+log("POST PROCESSING: Injecting JS symbols")
+
+
+def inject_v8_symbols(perf_dat_file):
+  output_file = perf_dat_file.with_suffix(".data.jitted")
+  cmd = [
+      "perf", "inject", "--jit", f"--input={perf_dat_file}",
+      f"--output={output_file}"
+  ]
+  try:
+    subprocess.run(cmd)
+    print(f"Processed: {output_file}")
+  except:
+    print(shlex.join(cmd))
+    return None
+  return output_file
+
+
+result = inject_v8_symbols(perf_data_file)
+if result is None:
+  print("No perf files were successfully processed"
+        " Check for errors or partial results in '{options.perf_data_dir}'")
+  exit(1)
+log(f"RESULTS in '{options.perf_data_dir}'")
+BYTES_TO_MIB = 1 / 1024 / 1024
+print(f"{result.name:67}{(result.stat().st_size*BYTES_TO_MIB):10.2f}MiB")
+
+log("PPROF")
+print(f"pprof -flame {result}")

tools/run-perf.sh

@@ -1,58 +0,0 @@
-#! /bin/sh
-#
-# Copyright 2016 the V8 project authors. All rights reserved.
-# Use of this source code is governed by a BSD-style license that can be
-# found in the LICENSE file.
-#
-
-########## Global variable definitions
-
-# Ensure that <your CPU clock> / $SAMPLE_EVERY_N_CYCLES < $MAXIMUM_SAMPLE_RATE.
-MAXIMUM_SAMPLE_RATE=10000000
-SAMPLE_EVERY_N_CYCLES=10000
-SAMPLE_RATE_CONFIG_FILE="/proc/sys/kernel/perf_event_max_sample_rate"
-KERNEL_MAP_CONFIG_FILE="/proc/sys/kernel/kptr_restrict"
-CALL_GRAPH_METHOD="fp"  # dwarf does not play nice with JITted objects.
-EVENT_TYPE=${EVENT_TYPE:=cycles:u}
-
-########## Usage
-
-usage() {
-cat << EOF
-usage: $0 <benchmark_command>
-
-Executes <benchmark_command> under observation by Linux perf.
-Sampling event is cycles in user space, call graphs are recorded.
-EOF
-}
-
-if [ $# -eq 0 ] || [ "$1" = "-h" ]  || [ "$1" = "--help" ] ; then
-  usage
-  exit 1
-fi
-
-########## Actual script execution
-
-ACTUAL_SAMPLE_RATE=$(cat $SAMPLE_RATE_CONFIG_FILE)
-if [ "$ACTUAL_SAMPLE_RATE" -lt "$MAXIMUM_SAMPLE_RATE" ] ; then
-  echo "Setting appropriate maximum sample rate..."
-  echo $MAXIMUM_SAMPLE_RATE | sudo tee $SAMPLE_RATE_CONFIG_FILE
-fi
-
-ACTUAL_KERNEL_MAP_RESTRICTION=$(cat $KERNEL_MAP_CONFIG_FILE)
-if [ "$ACTUAL_KERNEL_MAP_RESTRICTION" -ne "0" ] ; then
-  echo "Disabling kernel address map restriction..."
-  echo 0 | sudo tee $KERNEL_MAP_CONFIG_FILE
-fi
-
-# Extract the command being perfed, so that we can prepend arguments to the
-# arguments that the user supplied.
-COMMAND=$1
-shift 1
-
-echo "Running..."
-perf record -R \
-  -e $EVENT_TYPE \
-  -c $SAMPLE_EVERY_N_CYCLES \
-  --call-graph $CALL_GRAPH_METHOD \
-  -i "$COMMAND" --perf-basic-prof "$@"