Tools/Scripts/webkitpy/common/system/profiler.py

# Copyright (C) 2012 Google Inc. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are
# met:
#
#     * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
#     * Redistributions in binary form must reproduce the above
# copyright notice, this list of conditions and the following disclaimer
# in the documentation and/or other materials provided with the
# distribution.
#     * Neither the Google name nor the names of its
# contributors may be used to endorse or promote products derived from
# this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

from __future__ import print_function
import logging
import re
import sys

if sys.version_info < (3, 0):
    from itertools import ifilter  as filter

_log = logging.getLogger(__name__)


class ProfilerFactory(object):
    @classmethod
    def create_profiler(cls, host, executable_path, output_dir, profiler_name=None, identifier=None):
        profilers = cls.profilers_for_platform(host.platform)
        if not profilers:
            return None
        profiler_name = profiler_name or cls.default_profiler_name(host.platform)
        profiler_class = next(filter(lambda profiler: profiler.name == profiler_name, profilers), None)
        if not profiler_class:
            return None
        return profilers[0](host, executable_path, output_dir, identifier)

    @classmethod
    def default_profiler_name(cls, platform):
        profilers = cls.profilers_for_platform(platform)
        return profilers[0].name if profilers else None

    @classmethod
    def profilers_for_platform(cls, platform):
        # GooglePProf requires TCMalloc/google-perftools, but is available everywhere.
        profilers_by_os_name = {
            'mac': [IProfiler, Sample, GooglePProf],
            'linux': [Perf, GooglePProf],
            # Note: freebsd, win32 have no profilers defined yet, thus --profile will be ignored
            # by default, but a profiler can be selected with --profiler=PROFILER explicitly.
        }
        return profilers_by_os_name.get(platform.os_name, [])


class Profiler(object):
    # Used by ProfilerFactory to lookup a profiler from the --profiler=NAME option.
    name = None

    def __init__(self, host, executable_path, output_dir, identifier=None):
        self._host = host
        self._executable_path = executable_path
        self._output_dir = output_dir
        self._identifier = "test"
        self._host.filesystem.maybe_make_directory(self._output_dir)

    def adjusted_environment(self, env):
        return env

    def attach_to_pid(self, pid):
        pass

    def wrapper_arguments(self):
        return []

    def profile_after_exit(self):
        pass


class SingleFileOutputProfiler(Profiler):
    def __init__(self, host, executable_path, output_dir, output_suffix, identifier=None):
        super(SingleFileOutputProfiler, self).__init__(host, executable_path, output_dir, identifier)
        # FIXME: Currently all reports are kept as test.*, until we fix that, search up to 1000 names before giving up.
        self._output_path = self._host.workspace.find_unused_filename(self._output_dir, self._identifier, output_suffix, search_limit=1000)
        assert(self._output_path)


class GooglePProf(SingleFileOutputProfiler):
    name = 'pprof'

    def __init__(self, host, executable_path, output_dir, identifier=None):
        super(GooglePProf, self).__init__(host, executable_path, output_dir, "pprof", identifier)

    def adjusted_environment(self, env):
        env['CPUPROFILE'] = self._output_path
        return env

    def _first_ten_lines_of_profile(self, pprof_output):
        match = re.search("^Total:[^\n]*\n((?:[^\n]*\n){0,10})", pprof_output, re.MULTILINE)
        return match.group(1) if match else None

    def _pprof_path(self):
        # FIXME: We should have code to find the right google-pprof executable, some Googlers have
        # google-pprof installed as "pprof" on their machines for them.
        return '/usr/bin/google-pprof'

    def profile_after_exit(self):
        # google-pprof doesn't check its arguments, so we have to.
        if not (self._host.filesystem.exists(self._output_path)):
            print("Failed to gather profile, %s does not exist." % self._output_path)
            return

        pprof_args = [self._pprof_path(), '--text', self._executable_path, self._output_path]
        profile_text = self._host.executive.run_command(pprof_args)
        print("First 10 lines of pprof --text:")
        print(self._first_ten_lines_of_profile(profile_text))
        print("http://google-perftools.googlecode.com/svn/trunk/doc/cpuprofile.html documents output.")
        print()
        print("To interact with the the full profile, including produce graphs:")
        print(' '.join([self._pprof_path(), self._executable_path, self._output_path]))


class Perf(SingleFileOutputProfiler):
    name = 'perf'

    def __init__(self, host, executable_path, output_dir, identifier=None):
        super(Perf, self).__init__(host, executable_path, output_dir, "data", identifier)
        self._watched_pid = None
        self._wait_process = None

    def _perf_path(self):
        # FIXME: We may need to support finding the perf binary in other locations.
        return 'perf'

    def attach_to_pid(self, pid):
        # The passed-in pid here is the pid of the profiler. A wait process is launched here that
        # watches that pid and returns the same return code as the profiler process.
        self._watched_pid = pid
        self._wait_process = self._host.executive.popen(["wait", "%d" % pid], shell=True)

    def wrapper_arguments(self):
        return [self._perf_path(), "record", "-g", "--output", self._output_path]

    def _first_ten_lines_of_profile(self, perf_output):
        output_lines = re.finditer(r"^(?:( [^\n]*?)\s*\n)", perf_output, re.MULTILINE)
        prettified_lines = []
        for i in range(0, 10):
            line = next(output_lines, None)
            if not line:
                break
            prettified_lines.append(line.group(1))

        return prettified_lines

    def profile_after_exit(self):
        # Kill the watched process if it's still running.
        if self._wait_process.poll() is None:
            self._host.executive.kill_process(self._watched_pid)

        # Return early if the process produced non-zero exit code or is still running (if it couldn't be killed).
        exit_code = self._wait_process.poll()
        if exit_code != 0:
            print("'perf record' failed, ", end=' ')
            if exit_code:
                print("exit code was %i." % exit_code)
            else:
                print("the profiled process with pid %i is still running." % self._watched_pid)
            return

        perf_report_args = [self._perf_path(), 'report', '--call-graph', 'none', '--input', self._output_path]
        perf_report_output = self._host.executive.run_command(perf_report_args)
        print("First 10 lines of 'perf report --call-graph=none':")

        print(" ".join(perf_report_args))
        print("\n".join(self._first_ten_lines_of_profile(perf_report_output)))

        print("To view the full profile, run:")
        print(' '.join([self._perf_path(), 'report', '-i', self._output_path]))
        print()  # An extra line between tests looks nicer.


class Sample(SingleFileOutputProfiler):
    name = 'sample'

    def __init__(self, host, executable_path, output_dir, identifier=None):
        super(Sample, self).__init__(host, executable_path, output_dir, "txt", identifier)
        self._profiler_process = None

    def attach_to_pid(self, pid):
        cmd = ["sample", pid, "-mayDie", "-file", self._output_path]
        self._profiler_process = self._host.executive.popen(cmd)

    def profile_after_exit(self):
        self._profiler_process.wait()


class IProfiler(SingleFileOutputProfiler):
    name = 'iprofiler'

    def __init__(self, host, executable_path, output_dir, identifier=None):
        super(IProfiler, self).__init__(host, executable_path, output_dir, "dtps", identifier)
        self._profiler_process = None

    def attach_to_pid(self, pid):
        # FIXME: iprofiler requires us to pass the directory separately
        # from the basename of the file, with no control over the extension.
        fs = self._host.filesystem
        cmd = ["iprofiler", "-T", "0", "-timeprofiler", "-a", pid,
                "-d", fs.dirname(self._output_path), "-o", fs.splitext(fs.basename(self._output_path))[0]]
        # FIXME: Consider capturing instead of letting instruments spam to stderr directly.
        self._profiler_process = self._host.executive.popen(cmd)

    def profile_after_exit(self):
        # It seems like a nicer user experiance to wait on the profiler to exit to prevent
        # it from spewing to stderr at odd times.
        self._profiler_process.wait()