[hwasan] Offline symbolization script.

[eugenis]

Summary

A script to symbolize hwasan reports after the fact using unstripped
binaries. Supports stack-based reports. Requires llvm-symbolizer
(addr2line is not an option).

[eugenis]

removed leftover debug

[hctim]

compiler-rt/lib/hwasan/scripts/hwasan_symbolize

2

Should this file be hwasan_symbolize.py?

240

Shouldn't we be getting this from user PATH instead? For me , which llvm-symbolizer returns /usr/local/bin/llvm-symbolizer (which is the default install path for ninja install IIRC).

[eugenis]

look for llvm-symbolizer in PATH

[eugenis]

compiler-rt/lib/hwasan/scripts/hwasan_symbolize

2

it's installed in bin/, and we don't really want a .py extension there

240

fixed

[hctim]

LGTM with a couple comments

compiler-rt/lib/hwasan/scripts/hwasan_symbolize

120

I think a lot of .*?'s here can be replaced with \s* for stability:

(\s*)#(\d+)\s*(0x[0-9a-f]+)\s*\((.*)\+(0x[0-9a-f]+)\)

185

Nit: extra line

220

Needs to be updated to talk about PATH

248

We'll miss llvm-symbolizer in a user's PATH here.

[eugenis]

address comments

[eugenis]

compiler-rt/lib/hwasan/scripts/hwasan_symbolize

120

Not really. The first one needs to accept any prefix, like timestamp in adb logcat.
The second and the third have to be actual whitespace, we print it ourselves.

compiler-rt/lib/hwasan/CMakeLists.txt

add_sanitizer_rt_symbols(clang_rt.hwasan_cxx
       EXTRA hwasan.syms.extra)
     add_dependencies(hwasan clang_rt.hwasan-${arch}-symbols
                           clang_rt.hwasan_cxx-${arch}-symbols)
   endif()
 endforeach()
 
 add_compiler_rt_resource_file(hwasan_blacklist hwasan_blacklist.txt hwasan)
 
+add_subdirectory("scripts")
+
 # if(COMPILER_RT_INCLUDE_TESTS)
 #   add_subdirectory(tests)
 # endif()

compiler-rt/lib/hwasan/scripts/CMakeLists.txt

+add_compiler_rt_script(hwasan_symbolize)
+add_dependencies(hwasan hwasan_symbolize)

compiler-rt/lib/hwasan/scripts/hwasan_symbolize

+#!/usr/bin/env python
+#===- lib/hwasan/scripts/hwasan_symbolize ----------------------------------===#
+#
+# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+# See https:#llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+#===------------------------------------------------------------------------===#
+#
+# HWAddressSanitizer offline symbolization script.
+#
+#===------------------------------------------------------------------------===#
+import glob
+import os
+import re
+import sys
+import string
+import subprocess
+import argparse
+
+last_access_address = None
+last_access_tag = None
+
+class Symbolizer:
+  def __init__(self, path, binary_prefixes, paths_to_cut):
+    self.__pipe = None
+    self.__path = path
+    self.__binary_prefixes = binary_prefixes
+    self.__paths_to_cut = paths_to_cut
+    self.__log = False
+
+  def enable_logging(self, enable):
+    self.__log = enable
+
+  def __open_pipe(self):
+    if not self.__pipe:
+      self.__pipe = subprocess.Popen([self.__path, "-inlining", "-functions"],
+                                     stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+
+  class __EOF:
+    pass
+
+  def __write(self, s):
+    print >>self.__pipe.stdin, s
+    if self.__log:
+      print >>sys.stderr, ("#>>  |%s|" % (s,))
+
+  def __read(self):
+    s = self.__pipe.stdout.readline().rstrip()
+    if self.__log:
+      print >>sys.stderr, ("# << |%s|" % (s,))
+    if s == '':
+      raise Symbolizer.__EOF
+    return s
+
+  def __process_source_path(self, file_name):
+    for path_to_cut in self.__paths_to_cut:
+      file_name = re.sub(".*" + path_to_cut, "", file_name)
+    file_name = re.sub(".*hwasan_[a-z_]*.(cc|h):[0-9]*", "[hwasan_rtl]", file_name)
+    file_name = re.sub(".*asan_[a-z_]*.(cc|h):[0-9]*", "[asan_rtl]", file_name)
+    file_name = re.sub(".*crtstuff.c:0", "???:0", file_name)
+    return file_name
+
+  def __process_binary_name(self, name):
+    if name.startswith('/'):
+      name = name[1:]
+    for p in self.__binary_prefixes:
+      full_path = os.path.join(p, name)
+      if os.path.exists(full_path):
+        return full_path
+    # Try stripping extra path components as the last resort.
+    for p in self.__binary_prefixes:
+      full_path = os.path.join(p, os.path.basename(name))
+      if os.path.exists(full_path):
+        return full_path
+    print >>sys.stderr, "Could not find symbols for", name
+    return None
+
+  def iter_locals(self, binary, addr):
+    self.__open_pipe()
+    p = self.__pipe
+    binary = self.__process_binary_name(binary)
+    if not binary:
+      return
+    self.__write("FRAME %s %s" % (binary, addr))
+    try:
+      while True:
+        function_name = self.__read()
+        local_name = self.__read()
+        file_line = self.__read()
+        extra = self.__read().split()
+
+        file_line = self.__process_source_path(file_line)
+        offset = None if extra[0] == '??' else int(extra[0])
+        size = None if extra[1] == '??' else int(extra[1])
+        tag_offset = None if extra[2] == '??' else int(extra[2])
+        yield (function_name, file_line, local_name, offset, size, tag_offset)
+    except Symbolizer.__EOF:
+      pass
+
+  def iter_call_stack(self, binary, addr):
+    self.__open_pipe()
+    p = self.__pipe
+    binary = self.__process_binary_name(binary)
+    if not binary:
+      return
+    self.__write("CODE %s %s" % (binary, addr))
+    try:
+      while True:
+        function_name = self.__read()
+        file_line = self.__read()
+        file_line = self.__process_source_path(file_line)
+        yield (function_name, file_line)
+    except Symbolizer.__EOF:
+      pass
+
+def symbolize_line(line, symbolizer_path):
+  #0 0x7f6e35cf2e45  (/blah/foo.so+0x11fe45)
+  match = re.match(r'^(.*?)#([0-9]+)( *)(0x[0-9a-f]+) *\((.*)\+(0x[0-9a-f]+)\)', line, re.UNICODE)
+  if match:
+    frameno = match.group(2)
+    binary = match.group(5)
+    addr = int(match.group(6), 16)
+
+    frames = list(symbolizer.iter_call_stack(binary, addr))
+
+    if len(frames) > 0:
+      print "%s#%s%s%s in %s" % (match.group(1).encode('utf-8'), match.group(2).encode('utf-8'),
+                                 match.group(3).encode('utf-8'), frames[0][0], frames[0][1])
+      for i in range(1, len(frames)):
+        space1 = ' ' * match.end(1)
+        space2 = ' ' * (match.start(4) - match.end(1) - 2)
+        print "%s->%s%s in %s" % (space1, space2, frames[i][0], frames[i][1])
+    else:
+      print line.rstrip().encode('utf-8')
+  else:
+    print line.rstrip().encode('utf-8')
+
+def save_access_address(line):
+  global last_access_address, last_access_tag
+  match = re.match(r'^(.*?)HWAddressSanitizer: tag-mismatch on address (0x[0-9a-f]+) ', line, re.UNICODE)
+  if match:
+    last_access_address = int(match.group(2), 16)
+  match = re.match(r'^(.*?) of size [0-9]+ at 0x[0-9a-f]+ tags: ([0-9a-f]+)/[0-9a-f]+ \(ptr/mem\)', line, re.UNICODE)
+  if match:
+    last_access_tag = int(match.group(2), 16)
+
+def process_stack_history(line, symbolizer, ignore_tags=False):
+  if last_access_address is None or last_access_tag is None:
+    return
+  if re.match(r'Previously allocated frames:', line, re.UNICODE):
+    return True
+  pc_mask = (1 << 48) - 1
+  fp_mask = (1 << 20) - 1
+  # record_addr:0x1234ABCD record:0x1234ABCD (/path/to/binary+0x1234ABCD)
+  match = re.match(r'^(.*?)record_addr:(0x[0-9a-f]+) +record:(0x[0-9a-f]+) +\((.*)\+(0x[0-9a-f]+)\)', line, re.UNICODE)
+  if match:
+    record_addr = int(match.group(2), 16)
+    record = int(match.group(3), 16)
+    binary = match.group(4)
+    addr = int(match.group(5), 16)
+    base_tag = (record_addr >> 3) & 0xFF
+    fp = (record >> 48) << 4
+    pc = record & pc_mask
+
+    for local in symbolizer.iter_locals(binary, addr):
+      frame_offset = local[3]
+      size = local[4]
+      if frame_offset is None or size is None:
+        continue
+      obj_offset = (last_access_address - fp - frame_offset) & fp_mask
+      if obj_offset >= size:
+        continue
+      tag_offset = local[5]
+      if not ignore_tags and (tag_offset is None or base_tag ^ tag_offset != last_access_tag):
+        continue
+      print ''
+      print 'Potentially referenced stack object:'
+      print '  %d bytes inside variable "%s" in stack frame of function "%s"' % (obj_offset, local[2], local[0])
+      print '  at %s' % (local[1],)
+    return True
+  return False
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-d', action='store_true')
+parser.add_argument('-v', action='store_true')
+parser.add_argument('--ignore-tags', action='store_true')
+parser.add_argument('--symbols', action='append')
+parser.add_argument('--source', action='append')
+parser.add_argument('--symbolizer')
+parser.add_argument('args', nargs=argparse.REMAINDER)
+args = parser.parse_args()
+
+# Unstripped binaries location.
+binary_prefixes = args.symbols or []
+if not binary_prefixes:
+  if 'ANDROID_PRODUCT_OUT' in os.environ:
+    product_out = os.path.join(os.environ['ANDROID_PRODUCT_OUT'], 'symbols')
+    binary_prefixes.append(product_out)
+
+for p in binary_prefixes:
+  if not os.path.isdir(p):
+    print >>sys.stderr, "Symbols path does not exist or is not a directory:", p
+    sys.exit(1)
+
+# Source location.
+paths_to_cut = args.source or []
+if not paths_to_cut:
+  paths_to_cut.append(os.getcwd() + '/')
+  if 'ANDROID_BUILD_TOP' in os.environ:
+    paths_to_cut.append(os.environ['ANDROID_BUILD_TOP'] + '/')
+
+# llvm-symbolizer binary.
+# 1. --symbolizer flag
+# 2. environment variable
+# 3. unsuffixed binary in the current directory
+# 4. if inside Android platform, prebuilt binary at a known path
+# 5. first "llvm-symbolizer", then "llvm-symbolizer-$VER" with the
+#    highest available version in $PATH
+symbolizer_path = args.symbolizer
+if not symbolizer_path:
+  if 'LLVM_SYMBOLIZER_PATH' in os.environ:
+    symbolizer_path = os.environ['LLVM_SYMBOLIZER_PATH']
+  elif 'HWASAN_SYMBOLIZER_PATH' in os.environ:
+    symbolizer_path = os.environ['HWASAN_SYMBOLIZER_PATH']
+
+if not symbolizer_path:
+  s = os.path.join(os.path.dirname(sys.argv[0]), 'llvm-symbolizer')
+  if os.path.exists(s):
+    symbolizer_path = s
+
+if not symbolizer_path:
+  if 'ANDROID_BUILD_TOP' in os.environ:
+    s = os.path.join(os.environ['ANDROID_BUILD_TOP'], 'prebuilts/clang/host/linux-x86/llvm-binutils-stable/llvm-symbolizer')
+    if os.path.exists(s):
+      symbolizer_path = s
+
+if not symbolizer_path:
+  for path in os.environ["PATH"].split(os.pathsep):
+    p = os.path.join(path, 'llvm-symbolizer')
+    if os.path.exists(p):
+      symbolizer_path = p
+      break
+
+def extract_version(s):
+  idx = s.rfind('-')
+  if idx == -1:
+    return 0
+  x = float(s[idx + 1:])
+  return x
+
+if not symbolizer_path:
+  for path in os.environ["PATH"].split(os.pathsep):
+    candidates = glob.glob(os.path.join(path, 'llvm-symbolizer-*'))
+    if len(candidates) > 0:
+      candidates.sort(key = extract_version, reverse = True)
+      symbolizer_path = candidates[0]
+      break
+
+if not os.path.exists(symbolizer_path):
+  print >>sys.stderr, "Symbolizer path does not exist:", symbolizer_path
+  sys.exit(1)
+
+if args.v:
+  print "Looking for symbols in:"
+  for s in binary_prefixes:
+    print "  %s" % (s,)
+  print "Stripping source path prefixes:"
+  for s in paths_to_cut:
+    print "  %s" % (s,)
+  print "Using llvm-symbolizer binary in:\n  %s" % (symbolizer_path,)
+  print
+
+symbolizer = Symbolizer(symbolizer_path, binary_prefixes, paths_to_cut)
+symbolizer.enable_logging(args.d)
+
+for line in sys.stdin:
+  line = line.decode('utf-8')
+  save_access_address(line)
+  if process_stack_history(line, symbolizer, ignore_tags=args.ignore_tags):
+    continue
+  symbolize_line(line, symbolizer_path)