config/find_OOM_errors.py
author Eric Rahm <erahm@mozilla.com>
Thu, 10 Nov 2016 12:47:23 -0800
changeset 348770 e27050b02d69b88e19e0dcd151916f78259f9bb1
parent 129874 8f83edc05fa4d458d82668a95853c907544d0b8d
permissions -rw-r--r--
Bug 1313488 - Part 1: Convert XPCOM test TestDeadlockDetector to a gtest. r=froydnj This converts TestDeadlockDetector to a gtest. The logic for spawning off subprocesses is replaced with gtest's built-in death tests. On linux this will clone() the process and assert that the child process generates the appropriate assertion message. On OSX it will use fork(). In theory this should work on Windows as well buy spawning a new process but this test currently disabled there. MozReview-Commit-ID: 9Sl0hHBVGT3

#!/usr/bin/env python
# This Source Code Form is subject to the terms of the Mozilla Public
# License, v. 2.0. If a copy of the MPL was not distributed with this
# file, You can obtain one at http://mozilla.org/MPL/2.0/.
from __future__ import print_function

usage = """%prog: A test for OOM conditions in the shell.

%prog finds segfaults and other errors caused by incorrect handling of
allocation during OOM (out-of-memory) conditions.
"""

help = """Check for regressions only. This runs a set of files with a known
number of OOM errors (specified by REGRESSION_COUNT), and exits with a non-zero
result if more or less errors are found. See js/src/Makefile.in for invocation.
"""


import hashlib
import re
import shlex
import subprocess
import sys
import threading
import time

from optparse import OptionParser

#####################################################################
# Utility functions
#####################################################################
def run(args, stdin=None):
  class ThreadWorker(threading.Thread):
    def __init__(self, pipe):
      super(ThreadWorker, self).__init__()
      self.all = ""
      self.pipe = pipe
      self.setDaemon(True)

    def run(self):
      while True:
        line = self.pipe.readline()
        if line == '': break
        else:
          self.all += line

  try:
    if type(args) == str:
      args = shlex.split(args)

    args = [str(a) for a in args] # convert to strs

    stdin_pipe = subprocess.PIPE if stdin else None
    proc = subprocess.Popen(args, stdin=stdin_pipe, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
    if stdin_pipe:
      proc.stdin.write(stdin)
      proc.stdin.close()

    stdout_worker = ThreadWorker(proc.stdout)
    stderr_worker = ThreadWorker(proc.stderr)
    stdout_worker.start()
    stderr_worker.start()

    proc.wait()
    stdout_worker.join()
    stderr_worker.join()

  except KeyboardInterrupt as e:
    sys.exit(-1)

  stdout, stderr = stdout_worker.all, stderr_worker.all
  result = (stdout, stderr, proc.returncode)
  return result

def get_js_files():
  (out, err, exit) = run('find ../jit-test/tests -name "*.js"')
  if (err, exit) != ("", 0):
    sys.exit("Wrong directory, run from an objdir")
  return out.split()



#####################################################################
# Blacklisting
#####################################################################
def in_blacklist(sig):
  return sig in blacklist

def add_to_blacklist(sig):
  blacklist[sig] = blacklist.get(sig, 0)
  blacklist[sig] += 1

# How often is a particular lines important for this.
def count_lines():
  """Keep track of the amount of times individual lines occur, in order to
     prioritize the errors which occur most frequently."""
  counts = {}
  for string,count in blacklist.items():
    for line in string.split("\n"):
      counts[line] = counts.get(line, 0) + count

  lines = []
  for k,v in counts.items():
    lines.append("{0:6}: {1}".format(v, k))

  lines.sort()

  countlog = file("../OOM_count_log", "w")
  countlog.write("\n".join(lines))
  countlog.flush()
  countlog.close()


#####################################################################
# Output cleaning
#####################################################################
def clean_voutput(err):
  # Skip what we can't reproduce
  err = re.sub(r"^--\d+-- run: /usr/bin/dsymutil \"shell/js\"$", "", err, flags=re.MULTILINE)
  err = re.sub(r"^==\d+==", "", err, flags=re.MULTILINE)
  err = re.sub(r"^\*\*\d+\*\*", "", err, flags=re.MULTILINE)
  err = re.sub(r"^\s+by 0x[0-9A-Fa-f]+: ", "by: ", err, flags=re.MULTILINE)
  err = re.sub(r"^\s+at 0x[0-9A-Fa-f]+: ", "at: ", err, flags=re.MULTILINE)
  err = re.sub(r"(^\s+Address 0x)[0-9A-Fa-f]+( is not stack'd)", r"\1\2", err, flags=re.MULTILINE)
  err = re.sub(r"(^\s+Invalid write of size )\d+", r"\1x", err, flags=re.MULTILINE)
  err = re.sub(r"(^\s+Invalid read of size )\d+", r"\1x", err, flags=re.MULTILINE)
  err = re.sub(r"(^\s+Address 0x)[0-9A-Fa-f]+( is )\d+( bytes inside a block of size )[0-9,]+( free'd)", r"\1\2\3\4", err, flags=re.MULTILINE)

  # Skip the repeating bit due to the segfault
  lines = []
  for l in err.split('\n'):
    if l == " Process terminating with default action of signal 11 (SIGSEGV)":
      break
    lines.append(l)
  err = '\n'.join(lines)

  return err

def remove_failed_allocation_backtraces(err):
  lines = []

  add = True
  for l in err.split('\n'):

    # Set start and end conditions for including text
    if l == " The site of the failed allocation is:":
      add = False
    elif l[:2] not in ['by: ', 'at:']:
      add = True

    if add:
      lines.append(l)


  err = '\n'.join(lines)

  return err


def clean_output(err):
  err = re.sub(r"^js\(\d+,0x[0-9a-f]+\) malloc: \*\*\* error for object 0x[0-9a-f]+: pointer being freed was not allocated\n\*\*\* set a breakppoint in malloc_error_break to debug\n$", "pointer being freed was not allocated", err, flags=re.MULTILINE)

  return err


#####################################################################
# Consts, etc
#####################################################################

command_template = 'shell/js' \
                 + ' -m -j -p' \
                 + ' -e "const platform=\'darwin\'; const libdir=\'../jit-test/lib/\';"' \
                 + ' -f ../jit-test/lib/prolog.js' \
                 + ' -f {0}'


# Blacklists are things we don't want to see in our logs again (though we do
# want to count them when they happen). Whitelists we do want to see in our
# logs again, principally because the information we have isn't enough.

blacklist = {}
add_to_blacklist(r"('', '', 1)") # 1 means OOM if the shell hasn't launched yet.
add_to_blacklist(r"('', 'out of memory\n', 1)")

whitelist = set()
whitelist.add(r"('', 'out of memory\n', -11)") # -11 means OOM
whitelist.add(r"('', 'out of memory\nout of memory\n', -11)")



#####################################################################
# Program
#####################################################################

# Options
parser = OptionParser(usage=usage)
parser.add_option("-r", "--regression", action="store", metavar="REGRESSION_COUNT", help=help,
                  type="int", dest="regression", default=None)
                  
(OPTIONS, args) = parser.parse_args()


if OPTIONS.regression != None:
  # TODO: This should be expanded as we get a better hang of the OOM problems.
  # For now, we'll just check that the number of OOMs in one short file does not
  # increase.
  files = ["../jit-test/tests/arguments/args-createontrace.js"]
else:
  files = get_js_files()

  # Use a command-line arg to reduce the set of files
  if len (args):
    files = [f for f in files if f.find(args[0]) != -1]


if OPTIONS.regression == None:
  # Don't use a logfile, this is automated for tinderbox.
  log = file("../OOM_log", "w")


num_failures = 0
for f in files:

  # Run it once to establish boundaries
  command = (command_template + ' -O').format(f)
  out, err, exit = run(command)
  max = re.match(".*OOM max count: (\d+).*", out, flags=re.DOTALL).groups()[0]
  max = int(max)
  
  # OOMs don't recover well for the first 20 allocations or so.
  # TODO: revisit this.
  for i in range(20, max): 

    if OPTIONS.regression == None:
      print("Testing allocation {0}/{1} in {2}".format(i,max,f))
    else:
      sys.stdout.write('.') # something short for tinderbox, no space or \n

    command = (command_template + ' -A {0}').format(f, i)
    out, err, exit = run(command)

    # Success (5 is SM's exit code for controlled errors)
    if exit == 5 and err.find("out of memory") != -1:
      continue

    # Failure
    else:

      if OPTIONS.regression != None:
        # Just count them
        num_failures += 1
        continue

      #########################################################################
      # The regression tests ends above. The rest of this is for running  the
      # script manually.
      #########################################################################

      problem = str((out, err, exit))
      if in_blacklist(problem) and problem not in whitelist:
        add_to_blacklist(problem)
        continue

      add_to_blacklist(problem)


      # Get valgrind output for a good stack trace
      vcommand = "valgrind --dsymutil=yes -q --log-file=OOM_valgrind_log_file " + command
      run(vcommand)
      vout = file("OOM_valgrind_log_file").read()
      vout = clean_voutput(vout)
      sans_alloc_sites = remove_failed_allocation_backtraces(vout)

      # Don't print duplicate information
      if in_blacklist(sans_alloc_sites):
        add_to_blacklist(sans_alloc_sites)
        continue

      add_to_blacklist(sans_alloc_sites)

      log.write ("\n")
      log.write ("\n")
      log.write ("=========================================================================")
      log.write ("\n")
      log.write ("An allocation failure at\n\tallocation {0}/{1} in {2}\n\t"
                 "causes problems (detected using bug 624094)"
                 .format(i, max, f))
      log.write ("\n")
      log.write ("\n")

      log.write ("Command (from obj directory, using patch from bug 624094):\n  " + command)
      log.write ("\n")
      log.write ("\n")
      log.write ("stdout, stderr, exitcode:\n  " + problem)
      log.write ("\n")
      log.write ("\n")

      double_free = err.find("pointer being freed was not allocated") != -1
      oom_detected = err.find("out of memory") != -1
      multiple_oom_detected = err.find("out of memory\nout of memory") != -1
      segfault_detected = exit == -11

      log.write ("Diagnosis: ")
      log.write ("\n")
      if multiple_oom_detected:
        log.write ("  - Multiple OOMs reported")
        log.write ("\n")
      if segfault_detected:
        log.write ("  - segfault")
        log.write ("\n")
      if not oom_detected:
        log.write ("  - No OOM checking")
        log.write ("\n")
      if double_free:
        log.write ("  - Double free")
        log.write ("\n")

      log.write ("\n")

      log.write ("Valgrind info:\n" + vout)
      log.write ("\n")
      log.write ("\n")
      log.flush()

  if OPTIONS.regression == None:
    count_lines()

print()

# Do the actual regression check
if OPTIONS.regression != None:
  expected_num_failures = OPTIONS.regression

  if num_failures != expected_num_failures:

    print("TEST-UNEXPECTED-FAIL |", end='')
    if num_failures > expected_num_failures:
      print("More out-of-memory errors were found ({0}) than expected ({1}). "
            "This probably means an allocation site has been added without a "
            "NULL-check. If this is unavoidable, you can account for it by "
            "updating Makefile.in.".format(num_failures, expected_num_failures),
            end='')
    else:
      print("Congratulations, you have removed {0} out-of-memory error(s) "
            "({1} remain)! Please account for it by updating Makefile.in." 
            .format(expected_num_failures - num_failures, num_failures),
            end='')
    sys.exit(-1)
  else:
    print('TEST-PASS | find_OOM_errors | Found the expected number of OOM '
          'errors ({0})'.format(expected_num_failures))