tools/isolate_driver.py

   1 #!/usr/bin/env python
   2 # Copyright 2014 The Chromium Authors. All rights reserved.
   3 # Use of this source code is governed by a BSD-style license that can be
   4 # found in the LICENSE file.
   5
   6 """Adaptor script called through build/isolate.gypi.
   7
   8 Creates a wrapping .isolate which 'includes' the original one, that can be
   9 consumed by tools/swarming_client/isolate.py. Path variables are determined
  10 based on the current working directory. The relative_cwd in the .isolated file
  11 is determined based on the .isolate file that declare the 'command' variable to
  12 be used so the wrapping .isolate doesn't affect this value.
  13
  14 This script loads build.ninja and processes it to determine all the executables
  15 referenced by the isolated target. It adds them in the wrapping .isolate file.
  16
  17 WARNING: The target to use for build.ninja analysis is the base name of the
  18 .isolate file plus '_run'. For example, 'foo_test.isolate' would have the target
  19 'foo_test_run' analysed.
  20 """
  21
  22 import glob
  23 import json
  24 import logging
  25 import os
  26 import posixpath
  27 import StringIO
  28 import subprocess
  29 import sys
  30 import time
  31
  32 TOOLS_DIR = os.path.dirname(os.path.abspath(__file__))
  33 SWARMING_CLIENT_DIR = os.path.join(TOOLS_DIR, 'swarming_client')
  34 SRC_DIR = os.path.dirname(TOOLS_DIR)
  35
  36 sys.path.insert(0, SWARMING_CLIENT_DIR)
  37
  38 import isolate_format
  39
  40
  41 def load_ninja_recursively(build_dir, ninja_path, build_steps):
  42   """Crudely extracts all the subninja and build referenced in ninja_path.
  43
  44   In particular, it ignores rule and variable declarations. The goal is to be
  45   performant (well, as much as python can be performant) which is currently in
  46   the <200ms range for a complete chromium tree. As such the code is laid out
  47   for performance instead of readability.
  48   """
  49   logging.debug('Loading %s', ninja_path)
  50   try:
  51     with open(os.path.join(build_dir, ninja_path), 'rb') as f:
  52       line = None
  53       merge_line = ''
  54       subninja = []
  55       for line in f:
  56         line = line.rstrip()
  57         if not line:
  58           continue
  59
  60         if line[-1] == '$':
  61           # The next line needs to be merged in.
  62           merge_line += line[:-1]
  63           continue
  64
  65         if merge_line:
  66           line = merge_line + line
  67           merge_line = ''
  68
  69         statement = line[:line.find(' ')]
  70         if statement == 'build':
  71           # Save the dependency list as a raw string. Only the lines needed will
  72           # be processed with raw_build_to_deps(). This saves a good 70ms of
  73           # processing time.
  74           build_target, dependencies = line[6:].split(': ', 1)
  75           # Interestingly, trying to be smart and only saving the build steps
  76           # with the intended extensions ('', '.stamp', '.so') slows down
  77           # parsing even if 90% of the build rules can be skipped.
  78           # On Windows, a single step may generate two target, so split items
  79           # accordingly. It has only been seen for .exe/.exe.pdb combos.
  80           for i in build_target.strip().split():
  81             build_steps[i] = dependencies
  82         elif statement == 'subninja':
  83           subninja.append(line[9:])
  84   except IOError:
  85     print >> sys.stderr, 'Failed to open %s' % ninja_path
  86     raise
  87
  88   total = 1
  89   for rel_path in subninja:
  90     try:
  91       # Load each of the files referenced.
  92       # TODO(maruel): Skip the files known to not be needed. It saves an aweful
  93       # lot of processing time.
  94       total += load_ninja_recursively(build_dir, rel_path, build_steps)
  95     except IOError:
  96       print >> sys.stderr, '... as referenced by %s' % ninja_path
  97       raise
  98   return total
  99
 100
 101 def load_ninja(build_dir):
 102   """Loads the tree of .ninja files in build_dir."""
 103   build_steps = {}
 104   total = load_ninja_recursively(build_dir, 'build.ninja', build_steps)
 105   logging.info('Loaded %d ninja files, %d build steps', total, len(build_steps))
 106   return build_steps
 107
 108
 109 def using_blacklist(item):
 110   """Returns True if an item should be analyzed.
 111
 112   Ignores many rules that are assumed to not depend on a dynamic library. If
 113   the assumption doesn't hold true anymore for a file format, remove it from
 114   this list. This is simply an optimization.
 115   """
 116   # *.json is ignored below, *.isolated.gen.json is an exception, it is produced
 117   # by isolate_driver.py in 'test_isolation_mode==prepare'.
 118   if item.endswith('.isolated.gen.json'):
 119     return True
 120   IGNORED = (
 121     '.a', '.cc', '.css', '.dat', '.def', '.frag', '.h', '.html', '.isolate',
 122     '.js', '.json', '.manifest', '.o', '.obj', '.pak', '.png', '.pdb', '.py',
 123     '.strings', '.test', '.txt', '.vert',
 124   )
 125   # ninja files use native path format.
 126   ext = os.path.splitext(item)[1]
 127   if ext in IGNORED:
 128     return False
 129   # Special case Windows, keep .dll.lib but discard .lib.
 130   if item.endswith('.dll.lib'):
 131     return True
 132   if ext == '.lib':
 133     return False
 134   return item not in ('', '|', '||')
 135
 136
 137 def raw_build_to_deps(item):
 138   """Converts a raw ninja build statement into the list of interesting
 139   dependencies.
 140   """
 141   # TODO(maruel): Use a whitelist instead? .stamp, .so.TOC, .dylib.TOC,
 142   # .dll.lib, .exe and empty.
 143   # The first item is the build rule, e.g. 'link', 'cxx', 'phony', etc.
 144   return filter(using_blacklist, item.split(' ')[1:])
 145
 146
 147 def collect_deps(target, build_steps, dependencies_added, rules_seen):
 148   """Recursively adds all the interesting dependencies for |target|
 149   into |dependencies_added|.
 150   """
 151   if rules_seen is None:
 152     rules_seen = set()
 153   if target in rules_seen:
 154     # TODO(maruel): Figure out how it happens.
 155     logging.warning('Circular dependency for %s!', target)
 156     return
 157   rules_seen.add(target)
 158   try:
 159     dependencies = raw_build_to_deps(build_steps[target])
 160   except KeyError:
 161     logging.info('Failed to find a build step to generate: %s', target)
 162     return
 163   logging.debug('collect_deps(%s) -> %s', target, dependencies)
 164   for dependency in dependencies:
 165     dependencies_added.add(dependency)
 166     collect_deps(dependency, build_steps, dependencies_added, rules_seen)
 167
 168
 169 def post_process_deps(build_dir, dependencies):
 170   """Processes the dependency list with OS specific rules."""
 171   def filter_item(i):
 172     if i.endswith('.so.TOC'):
 173       # Remove only the suffix .TOC, not the .so!
 174       return i[:-4]
 175     if i.endswith('.dylib.TOC'):
 176       # Remove only the suffix .TOC, not the .dylib!
 177       return i[:-4]
 178     if i.endswith('.dll.lib'):
 179       # Remove only the suffix .lib, not the .dll!
 180       return i[:-4]
 181     return i
 182
 183   def is_exe(i):
 184     # This script is only for adding new binaries that are created as part of
 185     # the component build.
 186     ext = os.path.splitext(i)[1]
 187     # On POSIX, executables have no extension.
 188     if ext not in ('', '.dll', '.dylib', '.exe', '.nexe', '.so'):
 189       return False
 190
 191     # Check for execute access and strip directories. This gets rid of all the
 192     # phony rules.
 193     p = os.path.join(build_dir, i)
 194     return os.access(p, os.X_OK) and not os.path.isdir(p)
 195
 196   return filter(is_exe, map(filter_item, dependencies))
 197
 198
 199 def create_wrapper(args, isolate_index, isolated_index):
 200   """Creates a wrapper .isolate that add dynamic libs.
 201
 202   The original .isolate is not modified.
 203   """
 204   cwd = os.getcwd()
 205   isolate = args[isolate_index]
 206   # The code assumes the .isolate file is always specified path-less in cwd. Fix
 207   # if this assumption doesn't hold true.
 208   assert os.path.basename(isolate) == isolate, isolate
 209
 210   # This will look like ../out/Debug. This is based against cwd. Note that this
 211   # must equal the value provided as PRODUCT_DIR.
 212   build_dir = os.path.dirname(args[isolated_index])
 213
 214   # This will look like chrome/unit_tests.isolate. It is based against SRC_DIR.
 215   # It's used to calculate temp_isolate.
 216   src_isolate = os.path.relpath(os.path.join(cwd, isolate), SRC_DIR)
 217
 218   # The wrapping .isolate. This will look like
 219   # ../out/Debug/gen/chrome/unit_tests.isolate.
 220   temp_isolate = os.path.join(build_dir, 'gen', src_isolate)
 221   temp_isolate_dir = os.path.dirname(temp_isolate)
 222
 223   # Relative path between the new and old .isolate file.
 224   isolate_relpath = os.path.relpath(
 225       '.', temp_isolate_dir).replace(os.path.sep, '/')
 226
 227   # It's a big assumption here that the name of the isolate file matches the
 228   # primary target '_run'. Fix accordingly if this doesn't hold true, e.g.
 229   # complain to maruel@.
 230   target = isolate[:-len('.isolate')] + '_run'
 231   build_steps = load_ninja(build_dir)
 232   binary_deps = set()
 233   collect_deps(target, build_steps, binary_deps, None)
 234   binary_deps = post_process_deps(build_dir, binary_deps)
 235   logging.debug(
 236       'Binary dependencies:%s', ''.join('\n  ' + i for i in binary_deps))
 237
 238   # Now do actual wrapping .isolate.
 239   isolate_dict = {
 240     'includes': [
 241       posixpath.join(isolate_relpath, isolate),
 242     ],
 243     'variables': {
 244       # Will look like ['<(PRODUCT_DIR)/lib/flibuser_prefs.so'].
 245       'files': sorted(
 246           '<(PRODUCT_DIR)/%s' % i.replace(os.path.sep, '/')
 247           for i in binary_deps),
 248     },
 249   }
 250   if not os.path.isdir(temp_isolate_dir):
 251     os.makedirs(temp_isolate_dir)
 252   comment = (
 253       '# Warning: this file was AUTOGENERATED.\n'
 254       '# DO NO EDIT.\n')
 255   out = StringIO.StringIO()
 256   isolate_format.print_all(comment, isolate_dict, out)
 257   isolate_content = out.getvalue()
 258   with open(temp_isolate, 'wb') as f:
 259     f.write(isolate_content)
 260   logging.info('Added %d dynamic libs', len(binary_deps))
 261   logging.debug('%s', isolate_content)
 262   args[isolate_index] = temp_isolate
 263
 264
 265 def prepare_isolate_call(args, output):
 266   """Gathers all information required to run isolate.py later.
 267
 268   Dumps it as JSON to |output| file.
 269   """
 270   with open(output, 'wb') as f:
 271     json.dump({
 272       'args': args,
 273       'dir': os.getcwd(),
 274       'version': 1,
 275     }, f, indent=2, sort_keys=True)
 276
 277
 278 def main():
 279   logging.basicConfig(level=logging.ERROR, format='%(levelname)7s %(message)s')
 280   args = sys.argv[1:]
 281   mode = args[0] if args else None
 282   isolate = None
 283   isolated = None
 284   for i, arg in enumerate(args):
 285     if arg == '--isolate':
 286       isolate = i + 1
 287     if arg == '--isolated':
 288       isolated = i + 1
 289   if isolate is None or isolated is None or not mode:
 290     print >> sys.stderr, 'Internal failure'
 291     return 1
 292
 293   create_wrapper(args, isolate, isolated)
 294
 295   # In 'prepare' mode just collect all required information for postponed
 296   # isolated.py invocation later, store it in *.isolated.gen.json file.
 297   if mode == 'prepare':
 298     prepare_isolate_call(args[1:], args[isolated] + '.gen.json')
 299     return 0
 300
 301   swarming_client = os.path.join(SRC_DIR, 'tools', 'swarming_client')
 302   sys.stdout.flush()
 303   result = subprocess.call(
 304       [sys.executable, os.path.join(swarming_client, 'isolate.py')] + args)
 305   return result
 306
 307
 308 if __name__ == '__main__':
 309   sys.exit(main())