| OLD | NEW |
| (Empty) |
| 1 # Copyright 2014 The Chromium Authors. All rights reserved. | |
| 2 # Use of this source code is governed by a BSD-style license that can be | |
| 3 # found in the LICENSE file. | |
| 4 | |
| 5 """The ElfSymbolizer class for symbolizing Executable and Linkable Files. | |
| 6 | |
| 7 Adapted for Skia's use from | |
| 8 chromium/src/build/android/pylib/symbols/elf_symbolizer.py. | |
| 9 | |
| 10 Main changes: | |
| 11 -- Added prefix_to_remove param to remove path prefix from tree data. | |
| 12 """ | |
| 13 | |
| 14 import collections | |
| 15 import datetime | |
| 16 import logging | |
| 17 import multiprocessing | |
| 18 import os | |
| 19 import posixpath | |
| 20 import Queue | |
| 21 import re | |
| 22 import subprocess | |
| 23 import sys | |
| 24 import threading | |
| 25 | |
| 26 | |
| 27 # addr2line builds a possibly infinite memory cache that can exhaust | |
| 28 # the computer's memory if allowed to grow for too long. This constant | |
| 29 # controls how many lookups we do before restarting the process. 4000 | |
| 30 # gives near peak performance without extreme memory usage. | |
| 31 ADDR2LINE_RECYCLE_LIMIT = 4000 | |
| 32 | |
| 33 | |
| 34 class ELFSymbolizer(object): | |
| 35 """An uber-fast (multiprocessing, pipelined and asynchronous) ELF symbolizer. | |
| 36 | |
| 37 This class is a frontend for addr2line (part of GNU binutils), designed to | |
| 38 symbolize batches of large numbers of symbols for a given ELF file. It | |
| 39 supports sharding symbolization against many addr2line instances and | |
| 40 pipelining of multiple requests per each instance (in order to hide addr2line | |
| 41 internals and OS pipe latencies). | |
| 42 | |
| 43 The interface exhibited by this class is a very simple asynchronous interface, | |
| 44 which is based on the following three methods: | |
| 45 - SymbolizeAsync(): used to request (enqueue) resolution of a given address. | |
| 46 - The |callback| method: used to communicated back the symbol information. | |
| 47 - Join(): called to conclude the batch to gather the last outstanding results. | |
| 48 In essence, before the Join method returns, this class will have issued as | |
| 49 many callbacks as the number of SymbolizeAsync() calls. In this regard, note | |
| 50 that due to multiprocess sharding, callbacks can be delivered out of order. | |
| 51 | |
| 52 Some background about addr2line: | |
| 53 - it is invoked passing the elf path in the cmdline, piping the addresses in | |
| 54 its stdin and getting results on its stdout. | |
| 55 - it has pretty large response times for the first requests, but it | |
| 56 works very well in streaming mode once it has been warmed up. | |
| 57 - it doesn't scale by itself (on more cores). However, spawning multiple | |
| 58 instances at the same time on the same file is pretty efficient as they | |
| 59 keep hitting the pagecache and become mostly CPU bound. | |
| 60 - it might hang or crash, mostly for OOM. This class deals with both of these | |
| 61 problems. | |
| 62 | |
| 63 Despite the "scary" imports and the multi* words above, (almost) no multi- | |
| 64 threading/processing is involved from the python viewpoint. Concurrency | |
| 65 here is achieved by spawning several addr2line subprocesses and handling their | |
| 66 output pipes asynchronously. Therefore, all the code here (with the exception | |
| 67 of the Queue instance in Addr2Line) should be free from mind-blowing | |
| 68 thread-safety concerns. | |
| 69 | |
| 70 The multiprocess sharding works as follows: | |
| 71 The symbolizer tries to use the lowest number of addr2line instances as | |
| 72 possible (with respect of |max_concurrent_jobs|) and enqueue all the requests | |
| 73 in a single addr2line instance. For few symbols (i.e. dozens) sharding isn't | |
| 74 worth the startup cost. | |
| 75 The multiprocess logic kicks in as soon as the queues for the existing | |
| 76 instances grow. Specifically, once all the existing instances reach the | |
| 77 |max_queue_size| bound, a new addr2line instance is kicked in. | |
| 78 In the case of a very eager producer (i.e. all |max_concurrent_jobs| instances | |
| 79 have a backlog of |max_queue_size|), back-pressure is applied on the caller by | |
| 80 blocking the SymbolizeAsync method. | |
| 81 | |
| 82 This module has been deliberately designed to be dependency free (w.r.t. of | |
| 83 other modules in this project), to allow easy reuse in external projects. | |
| 84 """ | |
| 85 | |
| 86 def __init__(self, elf_file_path, addr2line_path, callback, inlines=False, | |
| 87 max_concurrent_jobs=None, addr2line_timeout=30, max_queue_size=50, | |
| 88 source_root_path=None, strip_base_path=None, prefix_to_remove=None): | |
| 89 """Args: | |
| 90 elf_file_path: path of the elf file to be symbolized. | |
| 91 addr2line_path: path of the toolchain's addr2line binary. | |
| 92 callback: a callback which will be invoked for each resolved symbol with | |
| 93 the two args (sym_info, callback_arg). The former is an instance of | |
| 94 |ELFSymbolInfo| and contains the symbol information. The latter is an | |
| 95 embedder-provided argument which is passed to SymbolizeAsync(). | |
| 96 inlines: when True, the ELFSymbolInfo will contain also the details about | |
| 97 the outer inlining functions. When False, only the innermost function | |
| 98 will be provided. | |
| 99 max_concurrent_jobs: Max number of addr2line instances spawned. | |
| 100 Parallelize responsibly, addr2line is a memory and I/O monster. | |
| 101 max_queue_size: Max number of outstanding requests per addr2line instance. | |
| 102 addr2line_timeout: Max time (in seconds) to wait for a addr2line response. | |
| 103 After the timeout, the instance will be considered hung and respawned. | |
| 104 source_root_path: In some toolchains only the name of the source file is | |
| 105 is output, without any path information; disambiguation searches | |
| 106 through the source directory specified by |source_root_path| argument | |
| 107 for files whose name matches, adding the full path information to the | |
| 108 output. For example, if the toolchain outputs "unicode.cc" and there | |
| 109 is a file called "unicode.cc" located under |source_root_path|/foo, | |
| 110 the tool will replace "unicode.cc" with | |
| 111 "|source_root_path|/foo/unicode.cc". If there are multiple files with | |
| 112 the same name, disambiguation will fail because the tool cannot | |
| 113 determine which of the files was the source of the symbol. | |
| 114 strip_base_path: Rebases the symbols source paths onto |source_root_path| | |
| 115 (i.e replace |strip_base_path| with |source_root_path). | |
| 116 prefix_to_remove: Removes the prefix from ElfSymbolInfo output. Skia added | |
| 117 """ | |
| 118 assert(os.path.isfile(addr2line_path)), 'Cannot find ' + addr2line_path | |
| 119 self.elf_file_path = elf_file_path | |
| 120 self.addr2line_path = addr2line_path | |
| 121 self.callback = callback | |
| 122 self.inlines = inlines | |
| 123 self.max_concurrent_jobs = (max_concurrent_jobs or | |
| 124 min(multiprocessing.cpu_count(), 4)) | |
| 125 self.max_queue_size = max_queue_size | |
| 126 self.addr2line_timeout = addr2line_timeout | |
| 127 self.requests_counter = 0 # For generating monotonic request IDs. | |
| 128 self._a2l_instances = [] # Up to |max_concurrent_jobs| _Addr2Line inst. | |
| 129 | |
| 130 # Skia addition: remove the given prefix from tree paths. | |
| 131 self.prefix_to_remove = prefix_to_remove | |
| 132 | |
| 133 # If necessary, create disambiguation lookup table | |
| 134 self.disambiguate = source_root_path is not None | |
| 135 self.disambiguation_table = {} | |
| 136 self.strip_base_path = strip_base_path | |
| 137 if(self.disambiguate): | |
| 138 self.source_root_path = os.path.abspath(source_root_path) | |
| 139 self._CreateDisambiguationTable() | |
| 140 | |
| 141 # Create one addr2line instance. More instances will be created on demand | |
| 142 # (up to |max_concurrent_jobs|) depending on the rate of the requests. | |
| 143 self._CreateNewA2LInstance() | |
| 144 | |
| 145 def SymbolizeAsync(self, addr, callback_arg=None): | |
| 146 """Requests symbolization of a given address. | |
| 147 | |
| 148 This method is not guaranteed to return immediately. It generally does, but | |
| 149 in some scenarios (e.g. all addr2line instances have full queues) it can | |
| 150 block to create back-pressure. | |
| 151 | |
| 152 Args: | |
| 153 addr: address to symbolize. | |
| 154 callback_arg: optional argument which will be passed to the |callback|.""" | |
| 155 assert(isinstance(addr, int)) | |
| 156 | |
| 157 # Process all the symbols that have been resolved in the meanwhile. | |
| 158 # Essentially, this drains all the addr2line(s) out queues. | |
| 159 for a2l_to_purge in self._a2l_instances: | |
| 160 a2l_to_purge.ProcessAllResolvedSymbolsInQueue() | |
| 161 a2l_to_purge.RecycleIfNecessary() | |
| 162 | |
| 163 # Find the best instance according to this logic: | |
| 164 # 1. Find an existing instance with the shortest queue. | |
| 165 # 2. If all of instances' queues are full, but there is room in the pool, | |
| 166 # (i.e. < |max_concurrent_jobs|) create a new instance. | |
| 167 # 3. If there were already |max_concurrent_jobs| instances and all of them | |
| 168 # had full queues, make back-pressure. | |
| 169 | |
| 170 # 1. | |
| 171 def _SortByQueueSizeAndReqID(a2l): | |
| 172 return (a2l.queue_size, a2l.first_request_id) | |
| 173 a2l = min(self._a2l_instances, key=_SortByQueueSizeAndReqID) | |
| 174 | |
| 175 # 2. | |
| 176 if (a2l.queue_size >= self.max_queue_size and | |
| 177 len(self._a2l_instances) < self.max_concurrent_jobs): | |
| 178 a2l = self._CreateNewA2LInstance() | |
| 179 | |
| 180 # 3. | |
| 181 if a2l.queue_size >= self.max_queue_size: | |
| 182 a2l.WaitForNextSymbolInQueue() | |
| 183 | |
| 184 a2l.EnqueueRequest(addr, callback_arg) | |
| 185 | |
| 186 def Join(self): | |
| 187 """Waits for all the outstanding requests to complete and terminates.""" | |
| 188 for a2l in self._a2l_instances: | |
| 189 a2l.WaitForIdle() | |
| 190 a2l.Terminate() | |
| 191 | |
| 192 def _CreateNewA2LInstance(self): | |
| 193 assert(len(self._a2l_instances) < self.max_concurrent_jobs) | |
| 194 a2l = ELFSymbolizer.Addr2Line(self) | |
| 195 self._a2l_instances.append(a2l) | |
| 196 return a2l | |
| 197 | |
| 198 def _CreateDisambiguationTable(self): | |
| 199 """ Non-unique file names will result in None entries""" | |
| 200 self.disambiguation_table = {} | |
| 201 | |
| 202 for root, _, filenames in os.walk(self.source_root_path): | |
| 203 for f in filenames: | |
| 204 self.disambiguation_table[f] = os.path.join(root, f) if (f not in | |
| 205 self.disambiguation_table) else None | |
| 206 | |
| 207 | |
| 208 class Addr2Line(object): | |
| 209 """A python wrapper around an addr2line instance. | |
| 210 | |
| 211 The communication with the addr2line process looks as follows: | |
| 212 [STDIN] [STDOUT] (from addr2line's viewpoint) | |
| 213 > f001111 | |
| 214 > f002222 | |
| 215 < Symbol::Name(foo, bar) for f001111 | |
| 216 < /path/to/source/file.c:line_number | |
| 217 > f003333 | |
| 218 < Symbol::Name2() for f002222 | |
| 219 < /path/to/source/file.c:line_number | |
| 220 < Symbol::Name3() for f003333 | |
| 221 < /path/to/source/file.c:line_number | |
| 222 """ | |
| 223 | |
| 224 SYM_ADDR_RE = re.compile(r'([^:]+):(\?|\d+).*') | |
| 225 | |
| 226 def __init__(self, symbolizer): | |
| 227 self._symbolizer = symbolizer | |
| 228 self._lib_file_name = posixpath.basename(symbolizer.elf_file_path) | |
| 229 | |
| 230 # The request queue (i.e. addresses pushed to addr2line's stdin and not | |
| 231 # yet retrieved on stdout) | |
| 232 self._request_queue = collections.deque() | |
| 233 | |
| 234 # This is essentially len(self._request_queue). It has been optimized to a | |
| 235 # separate field because turned out to be a perf hot-spot. | |
| 236 self.queue_size = 0 | |
| 237 | |
| 238 # Keep track of the number of symbols a process has processed to | |
| 239 # avoid a single process growing too big and using all the memory. | |
| 240 self._processed_symbols_count = 0 | |
| 241 | |
| 242 # Objects required to handle the addr2line subprocess. | |
| 243 self._proc = None # Subprocess.Popen(...) instance. | |
| 244 self._thread = None # Threading.thread instance. | |
| 245 self._out_queue = None # Queue.Queue instance (for buffering a2l stdout). | |
| 246 self._RestartAddr2LineProcess() | |
| 247 | |
| 248 def EnqueueRequest(self, addr, callback_arg): | |
| 249 """Pushes an address to addr2line's stdin (and keeps track of it).""" | |
| 250 self._symbolizer.requests_counter += 1 # For global "age" of requests. | |
| 251 req_idx = self._symbolizer.requests_counter | |
| 252 self._request_queue.append((addr, callback_arg, req_idx)) | |
| 253 self.queue_size += 1 | |
| 254 self._WriteToA2lStdin(addr) | |
| 255 | |
| 256 def WaitForIdle(self): | |
| 257 """Waits until all the pending requests have been symbolized.""" | |
| 258 while self.queue_size > 0: | |
| 259 self.WaitForNextSymbolInQueue() | |
| 260 | |
| 261 def WaitForNextSymbolInQueue(self): | |
| 262 """Waits for the next pending request to be symbolized.""" | |
| 263 if not self.queue_size: | |
| 264 return | |
| 265 | |
| 266 # This outer loop guards against a2l hanging (detecting stdout timeout). | |
| 267 while True: | |
| 268 start_time = datetime.datetime.now() | |
| 269 timeout = datetime.timedelta(seconds=self._symbolizer.addr2line_timeout) | |
| 270 | |
| 271 # The inner loop guards against a2l crashing (checking if it exited). | |
| 272 while (datetime.datetime.now() - start_time < timeout): | |
| 273 # poll() returns !None if the process exited. a2l should never exit. | |
| 274 if self._proc.poll(): | |
| 275 logging.warning('addr2line crashed, respawning (lib: %s).' % | |
| 276 self._lib_file_name) | |
| 277 self._RestartAddr2LineProcess() | |
| 278 # TODO(primiano): the best thing to do in this case would be | |
| 279 # shrinking the pool size as, very likely, addr2line is crashed | |
| 280 # due to low memory (and the respawned one will die again soon). | |
| 281 | |
| 282 try: | |
| 283 lines = self._out_queue.get(block=True, timeout=0.25) | |
| 284 except Queue.Empty: | |
| 285 # On timeout (1/4 s.) repeat the inner loop and check if either the | |
| 286 # addr2line process did crash or we waited its output for too long. | |
| 287 continue | |
| 288 | |
| 289 # In nominal conditions, we get straight to this point. | |
| 290 self._ProcessSymbolOutput(lines) | |
| 291 return | |
| 292 | |
| 293 # If this point is reached, we waited more than |addr2line_timeout|. | |
| 294 logging.warning('Hung addr2line process, respawning (lib: %s).' % | |
| 295 self._lib_file_name) | |
| 296 self._RestartAddr2LineProcess() | |
| 297 | |
| 298 def ProcessAllResolvedSymbolsInQueue(self): | |
| 299 """Consumes all the addr2line output lines produced (without blocking).""" | |
| 300 if not self.queue_size: | |
| 301 return | |
| 302 while True: | |
| 303 try: | |
| 304 lines = self._out_queue.get_nowait() | |
| 305 except Queue.Empty: | |
| 306 break | |
| 307 self._ProcessSymbolOutput(lines) | |
| 308 | |
| 309 def RecycleIfNecessary(self): | |
| 310 """Restarts the process if it has been used for too long. | |
| 311 | |
| 312 A long running addr2line process will consume excessive amounts | |
| 313 of memory without any gain in performance.""" | |
| 314 if self._processed_symbols_count >= ADDR2LINE_RECYCLE_LIMIT: | |
| 315 self._RestartAddr2LineProcess() | |
| 316 | |
| 317 | |
| 318 def Terminate(self): | |
| 319 """Kills the underlying addr2line process. | |
| 320 | |
| 321 The poller |_thread| will terminate as well due to the broken pipe.""" | |
| 322 try: | |
| 323 self._proc.kill() | |
| 324 self._proc.communicate() # Essentially wait() without risking deadlock. | |
| 325 except Exception: # An exception while terminating? How interesting. | |
| 326 pass | |
| 327 self._proc = None | |
| 328 | |
| 329 def _WriteToA2lStdin(self, addr): | |
| 330 self._proc.stdin.write('%s\n' % hex(addr)) | |
| 331 if self._symbolizer.inlines: | |
| 332 # In the case of inlines we output an extra blank line, which causes | |
| 333 # addr2line to emit a (??,??:0) tuple that we use as a boundary marker. | |
| 334 self._proc.stdin.write('\n') | |
| 335 self._proc.stdin.flush() | |
| 336 | |
| 337 def _ProcessSymbolOutput(self, lines): | |
| 338 """Parses an addr2line symbol output and triggers the client callback.""" | |
| 339 (_, callback_arg, _) = self._request_queue.popleft() | |
| 340 self.queue_size -= 1 | |
| 341 | |
| 342 innermost_sym_info = None | |
| 343 sym_info = None | |
| 344 for (line1, line2) in lines: | |
| 345 prev_sym_info = sym_info | |
| 346 name = line1 if not line1.startswith('?') else None | |
| 347 source_path = None | |
| 348 source_line = None | |
| 349 m = ELFSymbolizer.Addr2Line.SYM_ADDR_RE.match(line2) | |
| 350 if m: | |
| 351 if not m.group(1).startswith('?'): | |
| 352 source_path = m.group(1) | |
| 353 if not m.group(2).startswith('?'): | |
| 354 source_line = int(m.group(2)) | |
| 355 else: | |
| 356 logging.warning('Got invalid symbol path from addr2line: %s' % line2) | |
| 357 | |
| 358 # In case disambiguation is on, and needed | |
| 359 was_ambiguous = False | |
| 360 disambiguated = False | |
| 361 if self._symbolizer.disambiguate: | |
| 362 if source_path and not posixpath.isabs(source_path): | |
| 363 path = self._symbolizer.disambiguation_table.get(source_path) | |
| 364 was_ambiguous = True | |
| 365 disambiguated = path is not None | |
| 366 source_path = path if disambiguated else source_path | |
| 367 | |
| 368 # Use absolute paths (so that paths are consistent, as disambiguation | |
| 369 # uses absolute paths) | |
| 370 if source_path and not was_ambiguous: | |
| 371 source_path = os.path.abspath(source_path) | |
| 372 | |
| 373 if source_path and self._symbolizer.strip_base_path: | |
| 374 # Strip the base path | |
| 375 source_path = re.sub('^' + self._symbolizer.strip_base_path, | |
| 376 self._symbolizer.source_root_path or '', source_path) | |
| 377 | |
| 378 sym_info = ELFSymbolInfo(name, source_path, source_line, was_ambiguous, | |
| 379 disambiguated, | |
| 380 self._symbolizer.prefix_to_remove) | |
| 381 if prev_sym_info: | |
| 382 prev_sym_info.inlined_by = sym_info | |
| 383 if not innermost_sym_info: | |
| 384 innermost_sym_info = sym_info | |
| 385 | |
| 386 self._processed_symbols_count += 1 | |
| 387 self._symbolizer.callback(innermost_sym_info, callback_arg) | |
| 388 | |
| 389 def _RestartAddr2LineProcess(self): | |
| 390 if self._proc: | |
| 391 self.Terminate() | |
| 392 | |
| 393 # The only reason of existence of this Queue (and the corresponding | |
| 394 # Thread below) is the lack of a subprocess.stdout.poll_avail_lines(). | |
| 395 # Essentially this is a pipe able to extract a couple of lines atomically. | |
| 396 self._out_queue = Queue.Queue() | |
| 397 | |
| 398 # Start the underlying addr2line process in line buffered mode. | |
| 399 | |
| 400 cmd = [self._symbolizer.addr2line_path, '--functions', '--demangle', | |
| 401 '--exe=' + self._symbolizer.elf_file_path] | |
| 402 if self._symbolizer.inlines: | |
| 403 cmd += ['--inlines'] | |
| 404 self._proc = subprocess.Popen(cmd, bufsize=1, stdout=subprocess.PIPE, | |
| 405 stdin=subprocess.PIPE, stderr=sys.stderr, close_fds=True) | |
| 406 | |
| 407 # Start the poller thread, which simply moves atomically the lines read | |
| 408 # from the addr2line's stdout to the |_out_queue|. | |
| 409 self._thread = threading.Thread( | |
| 410 target=ELFSymbolizer.Addr2Line.StdoutReaderThread, | |
| 411 args=(self._proc.stdout, self._out_queue, self._symbolizer.inlines)) | |
| 412 self._thread.daemon = True # Don't prevent early process exit. | |
| 413 self._thread.start() | |
| 414 | |
| 415 self._processed_symbols_count = 0 | |
| 416 | |
| 417 # Replay the pending requests on the new process (only for the case | |
| 418 # of a hung addr2line timing out during the game). | |
| 419 for (addr, _, _) in self._request_queue: | |
| 420 self._WriteToA2lStdin(addr) | |
| 421 | |
| 422 @staticmethod | |
| 423 def StdoutReaderThread(process_pipe, queue, inlines): | |
| 424 """The poller thread fn, which moves the addr2line stdout to the |queue|. | |
| 425 | |
| 426 This is the only piece of code not running on the main thread. It merely | |
| 427 writes to a Queue, which is thread-safe. In the case of inlines, it | |
| 428 detects the ??,??:0 marker and sends the lines atomically, such that the | |
| 429 main thread always receives all the lines corresponding to one symbol in | |
| 430 one shot.""" | |
| 431 try: | |
| 432 lines_for_one_symbol = [] | |
| 433 while True: | |
| 434 line1 = process_pipe.readline().rstrip('\r\n') | |
| 435 line2 = process_pipe.readline().rstrip('\r\n') | |
| 436 if not line1 or not line2: | |
| 437 break | |
| 438 inline_has_more_lines = inlines and (len(lines_for_one_symbol) == 0 or | |
| 439 (line1 != '??' and line2 != '??:0')) | |
| 440 if not inlines or inline_has_more_lines: | |
| 441 lines_for_one_symbol += [(line1, line2)] | |
| 442 if inline_has_more_lines: | |
| 443 continue | |
| 444 queue.put(lines_for_one_symbol) | |
| 445 lines_for_one_symbol = [] | |
| 446 process_pipe.close() | |
| 447 | |
| 448 # Every addr2line processes will die at some point, please die silently. | |
| 449 except (IOError, OSError): | |
| 450 pass | |
| 451 | |
| 452 @property | |
| 453 def first_request_id(self): | |
| 454 """Returns the request_id of the oldest pending request in the queue.""" | |
| 455 return self._request_queue[0][2] if self._request_queue else 0 | |
| 456 | |
| 457 | |
| 458 class ELFSymbolInfo(object): | |
| 459 """The result of the symbolization passed as first arg. of each callback.""" | |
| 460 | |
| 461 def __init__(self, name, source_path, source_line, was_ambiguous=False, | |
| 462 disambiguated=False, prefix_to_remove=None): | |
| 463 """All the fields here can be None (if addr2line replies with '??').""" | |
| 464 self.name = name | |
| 465 if source_path and source_path.startswith(prefix_to_remove): | |
| 466 source_path = source_path[len(prefix_to_remove) : ] | |
| 467 self.source_path = source_path | |
| 468 self.source_line = source_line | |
| 469 # In the case of |inlines|=True, the |inlined_by| points to the outer | |
| 470 # function inlining the current one (and so on, to form a chain). | |
| 471 self.inlined_by = None | |
| 472 self.disambiguated = disambiguated | |
| 473 self.was_ambiguous = was_ambiguous | |
| 474 | |
| 475 def __str__(self): | |
| 476 return '%s [%s:%d]' % ( | |
| 477 self.name or '??', self.source_path or '??', self.source_line or 0) | |
| OLD | NEW |