subprocession: limit fd leaks

vcsserver/subprocessio.py

@@ -1,516 +1,523 b''
 """
 Module provides a class allowing to wrap communication over subprocess.Popen
 input, output, error streams into a meaningfull, non-blocking, concurrent
 stream processor exposing the output data as an iterator fitting to be a
 return value passed by a WSGI applicaiton to a WSGI server per PEP 3333.
 
 Copyright (c) 2011  Daniel Dotsenko <dotsa[at]hotmail.com>
 
 This file is part of git_http_backend.py Project.
 
 git_http_backend.py Project is free software: you can redistribute it and/or
 modify it under the terms of the GNU Lesser General Public License as
 published by the Free Software Foundation, either version 2.1 of the License,
 or (at your option) any later version.
 
 git_http_backend.py Project is distributed in the hope that it will be useful,
 but WITHOUT ANY WARRANTY; without even the implied warranty of
 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
 GNU Lesser General Public License for more details.
 
 You should have received a copy of the GNU Lesser General Public License
 along with git_http_backend.py Project.
 If not, see <http://www.gnu.org/licenses/>.
 """
 import os
 import logging
 import subprocess32 as subprocess
 from collections import deque
 from threading import Event, Thread
 
 log = logging.getLogger(__name__)
 
 
 class StreamFeeder(Thread):
     """
     Normal writing into pipe-like is blocking once the buffer is filled.
     This thread allows a thread to seep data from a file-like into a pipe
     without blocking the main thread.
     We close inpipe once the end of the source stream is reached.
     """
 
     def __init__(self, source):
         super(StreamFeeder, self).__init__()
         self.daemon = True
         filelike = False
         self.bytes = bytes()
         if type(source) in (type(''), bytes, bytearray):  # string-like
             self.bytes = bytes(source)
         else:  # can be either file pointer or file-like
             if type(source) in (int, long):  # file pointer it is
                 # converting file descriptor (int) stdin into file-like
                 try:
                     source = os.fdopen(source, 'rb', 16384)
                 except Exception:
                     pass
             # let's see if source is file-like by now
             try:
                 filelike = source.read
             except Exception:
                 pass
         if not filelike and not self.bytes:
             raise TypeError("StreamFeeder's source object must be a readable "
                             "file-like, a file descriptor, or a string-like.")
         self.source = source
         self.readiface, self.writeiface = os.pipe()
 
     def run(self):
         t = self.writeiface
-        if self.bytes:
-            os.write(t, self.bytes)
-        else:
-            s = self.source
-            b = s.read(4096)
-            while b:
-                os.write(t, b)
+        try:
+            if self.bytes:
+                os.write(t, self.bytes)
+            else:
+                s = self.source
                 b = s.read(4096)
-        os.close(t)
+                while b:
+                    os.write(t, b)
+                    b = s.read(4096)
+        finally:
+            os.close(t)
 
     @property
     def output(self):
         return self.readiface
 
 
 class InputStreamChunker(Thread):
     def __init__(self, source, target, buffer_size, chunk_size):
 
         super(InputStreamChunker, self).__init__()
 
         self.daemon = True  # die die die.
 
         self.source = source
         self.target = target
         self.chunk_count_max = int(buffer_size / chunk_size) + 1
         self.chunk_size = chunk_size
 
         self.data_added = Event()
         self.data_added.clear()
 
         self.keep_reading = Event()
         self.keep_reading.set()
 
         self.EOF = Event()
         self.EOF.clear()
 
         self.go = Event()
         self.go.set()
 
     def stop(self):
         self.go.clear()
         self.EOF.set()
         try:
             # this is not proper, but is done to force the reader thread let
             # go of the input because, if successful, .close() will send EOF
             # down the pipe.
             self.source.close()
         except:
             pass
 
     def run(self):
         s = self.source
         t = self.target
         cs = self.chunk_size
         chunk_count_max = self.chunk_count_max
         keep_reading = self.keep_reading
         da = self.data_added
         go = self.go
 
         try:
             b = s.read(cs)
         except ValueError:
             b = ''
 
         timeout_input = 20
         while b and go.is_set():
             if len(t) > chunk_count_max:
                 keep_reading.clear()
                 keep_reading.wait(timeout_input)
                 if len(t) > chunk_count_max + timeout_input:
                     log.error("Timed out while waiting for input from subprocess.")
                     os._exit(-1)  # this will cause the worker to recycle itself
 
             t.append(b)
             da.set()
 
             try:
                 b = s.read(cs)
             except ValueError:
                 b = ''
 
         self.EOF.set()
         da.set()  # for cases when done but there was no input.
 
 
 class BufferedGenerator(object):
     """
     Class behaves as a non-blocking, buffered pipe reader.
     Reads chunks of data (through a thread)
     from a blocking pipe, and attaches these to an array (Deque) of chunks.
     Reading is halted in the thread when max chunks is internally buffered.
     The .next() may operate in blocking or non-blocking fashion by yielding
     '' if no data is ready
     to be sent or by not returning until there is some data to send
     When we get EOF from underlying source pipe we raise the marker to raise
     StopIteration after the last chunk of data is yielded.
     """
 
     def __init__(self, source, buffer_size=65536, chunk_size=4096,
                  starting_values=None, bottomless=False):
         starting_values = starting_values or []
 
         if bottomless:
             maxlen = int(buffer_size / chunk_size)
         else:
             maxlen = None
 
         self.data = deque(starting_values, maxlen)
         self.worker = InputStreamChunker(source, self.data, buffer_size,
                                          chunk_size)
         if starting_values:
             self.worker.data_added.set()
         self.worker.start()
 
     ####################
     # Generator's methods
     ####################
 
     def __iter__(self):
         return self
 
     def next(self):
         while not len(self.data) and not self.worker.EOF.is_set():
             self.worker.data_added.clear()
             self.worker.data_added.wait(0.2)
         if len(self.data):
             self.worker.keep_reading.set()
             return bytes(self.data.popleft())
         elif self.worker.EOF.is_set():
             raise StopIteration
 
     def throw(self, exc_type, value=None, traceback=None):
         if not self.worker.EOF.is_set():
             raise exc_type(value)
 
     def start(self):
         self.worker.start()
 
     def stop(self):
         self.worker.stop()
 
     def close(self):
         try:
             self.worker.stop()
             self.throw(GeneratorExit)
         except (GeneratorExit, StopIteration):
             pass
 
     def __del__(self):
         self.close()
 
     ####################
     # Threaded reader's infrastructure.
     ####################
     @property
     def input(self):
         return self.worker.w
 
     @property
     def data_added_event(self):
         return self.worker.data_added
 
     @property
     def data_added(self):
         return self.worker.data_added.is_set()
 
     @property
     def reading_paused(self):
         return not self.worker.keep_reading.is_set()
 
     @property
     def done_reading_event(self):
         """
         Done_reding does not mean that the iterator's buffer is empty.
         Iterator might have done reading from underlying source, but the read
         chunks might still be available for serving through .next() method.
 
         :returns: An Event class instance.
         """
         return self.worker.EOF
 
     @property
     def done_reading(self):
         """
         Done_reding does not mean that the iterator's buffer is empty.
         Iterator might have done reading from underlying source, but the read
         chunks might still be available for serving through .next() method.
 
         :returns: An Bool value.
         """
         return self.worker.EOF.is_set()
 
     @property
     def length(self):
         """
         returns int.
 
         This is the lenght of the que of chunks, not the length of
         the combined contents in those chunks.
 
         __len__() cannot be meaningfully implemented because this
         reader is just flying throuh a bottomless pit content and
         can only know the lenght of what it already saw.
 
         If __len__() on WSGI server per PEP 3333 returns a value,
         the responce's length will be set to that. In order not to
         confuse WSGI PEP3333 servers, we will not implement __len__
         at all.
         """
         return len(self.data)
 
     def prepend(self, x):
         self.data.appendleft(x)
 
     def append(self, x):
         self.data.append(x)
 
     def extend(self, o):
         self.data.extend(o)
 
     def __getitem__(self, i):
         return self.data[i]
 
 
 class SubprocessIOChunker(object):
     """
     Processor class wrapping handling of subprocess IO.
 
     .. important::
 
        Watch out for the method `__del__` on this class. If this object
        is deleted, it will kill the subprocess, so avoid to
        return the `output` attribute or usage of it like in the following
        example::
 
           # `args` expected to run a program that produces a lot of output
           output = ''.join(SubprocessIOChunker(
              args, shell=False, inputstream=inputstream, env=environ).output)
 
           # `output` will not contain all the data, because the __del__ method
           # has already killed the subprocess in this case before all output
           # has been consumed.
 
 
 
     In a way, this is a "communicate()" replacement with a twist.
 
     - We are multithreaded. Writing in and reading out, err are all sep threads.
     - We support concurrent (in and out) stream processing.
     - The output is not a stream. It's a queue of read string (bytes, not unicode)
       chunks. The object behaves as an iterable. You can "for chunk in obj:" us.
     - We are non-blocking in more respects than communicate()
       (reading from subprocess out pauses when internal buffer is full, but
        does not block the parent calling code. On the flip side, reading from
        slow-yielding subprocess may block the iteration until data shows up. This
        does not block the parallel inpipe reading occurring parallel thread.)
 
     The purpose of the object is to allow us to wrap subprocess interactions into
     and interable that can be passed to a WSGI server as the application's return
     value. Because of stream-processing-ability, WSGI does not have to read ALL
     of the subprocess's output and buffer it, before handing it to WSGI server for
     HTTP response. Instead, the class initializer reads just a bit of the stream
     to figure out if error ocurred or likely to occur and if not, just hands the
     further iteration over subprocess output to the server for completion of HTTP
     response.
 
     The real or perceived subprocess error is trapped and raised as one of
     EnvironmentError family of exceptions
 
     Example usage:
     #    try:
     #        answer = SubprocessIOChunker(
     #            cmd,
     #            input,
     #            buffer_size = 65536,
     #            chunk_size = 4096
     #            )
     #    except (EnvironmentError) as e:
     #        print str(e)
     #        raise e
     #
     #    return answer
 
 
     """
 
     # TODO: johbo: This is used to make sure that the open end of the PIPE
     # is closed in the end. It would be way better to wrap this into an
     # object, so that it is closed automatically once it is consumed or
     # something similar.
     _close_input_fd = None
 
     _closed = False
 
     def __init__(self, cmd, inputstream=None, buffer_size=65536,
                  chunk_size=4096, starting_values=None, fail_on_stderr=True,
                  fail_on_return_code=True, **kwargs):
         """
         Initializes SubprocessIOChunker
 
         :param cmd: A Subprocess.Popen style "cmd". Can be string or array of strings
         :param inputstream: (Default: None) A file-like, string, or file pointer.
         :param buffer_size: (Default: 65536) A size of total buffer per stream in bytes.
         :param chunk_size: (Default: 4096) A max size of a chunk. Actual chunk may be smaller.
         :param starting_values: (Default: []) An array of strings to put in front of output que.
         :param fail_on_stderr: (Default: True) Whether to raise an exception in
                                case something is written to stderr.
         :param fail_on_return_code: (Default: True) Whether to raise an
                                     exception if the return code is not 0.
         """
 
         starting_values = starting_values or []
         if inputstream:
             input_streamer = StreamFeeder(inputstream)
             input_streamer.start()
             inputstream = input_streamer.output
             self._close_input_fd = inputstream
 
         self._fail_on_stderr = fail_on_stderr
         self._fail_on_return_code = fail_on_return_code
 
         _shell = kwargs.get('shell', True)
         kwargs['shell'] = _shell
 
         _p = subprocess.Popen(cmd, bufsize=-1,
                               stdin=inputstream,
                               stdout=subprocess.PIPE,
                               stderr=subprocess.PIPE,
                               **kwargs)
 
         bg_out = BufferedGenerator(_p.stdout, buffer_size, chunk_size,
                                    starting_values)
         bg_err = BufferedGenerator(_p.stderr, 16000, 1, bottomless=True)
 
         while not bg_out.done_reading and not bg_out.reading_paused and not bg_err.length:
             # doing this until we reach either end of file, or end of buffer.
             bg_out.data_added_event.wait(1)
             bg_out.data_added_event.clear()
 
         # at this point it's still ambiguous if we are done reading or just full buffer.
         # Either way, if error (returned by ended process, or implied based on
         # presence of stuff in stderr output) we error out.
         # Else, we are happy.
         _returncode = _p.poll()
 
         if ((_returncode and fail_on_return_code) or
                 (fail_on_stderr and _returncode is None and bg_err.length)):
             try:
                 _p.terminate()
             except Exception:
                 pass
             bg_out.stop()
             bg_err.stop()
             if fail_on_stderr:
                 err = ''.join(bg_err)
                 raise EnvironmentError(
                     "Subprocess exited due to an error:\n" + err)
             if _returncode and fail_on_return_code:
                 err = ''.join(bg_err)
                 if not err:
                     # maybe get empty stderr, try stdout instead
                     # in many cases git reports the errors on stdout too
                     err = ''.join(bg_out)
                 raise EnvironmentError(
                     "Subprocess exited with non 0 ret code:%s: stderr:%s" % (
                         _returncode, err))
 
         self.process = _p
         self.output = bg_out
         self.error = bg_err
+        self.inputstream = inputstream
 
     def __iter__(self):
         return self
 
     def next(self):
         # Note: mikhail: We need to be sure that we are checking the return
         # code after the stdout stream is closed. Some processes, e.g. git
         # are doing some magic in between closing stdout and terminating the
         # process and, as a result,  we are not getting return code on "slow"
         # systems.
         result = None
         stop_iteration = None
         try:
             result = self.output.next()
         except StopIteration as e:
             stop_iteration = e
 
         if self.process.poll() and self._fail_on_return_code:
             err = '%s' % ''.join(self.error)
             raise EnvironmentError(
                 "Subprocess exited due to an error:\n" + err)
 
         if stop_iteration:
             raise stop_iteration
         return result
 
     def throw(self, type, value=None, traceback=None):
         if self.output.length or not self.output.done_reading:
             raise type(value)
 
     def close(self):
         if self._closed:
             return
         self._closed = True
         try:
             self.process.terminate()
         except:
             pass
         if self._close_input_fd:
             os.close(self._close_input_fd)
         try:
             self.output.close()
         except:
             pass
         try:
             self.error.close()
         except:
             pass
+        try:
+            os.close(self.inputstream)
+        except:
+            pass
 
     def __del__(self):
         self.close()
 
 
 def run_command(arguments, env=None):
     """
     Run the specified command and return the stdout.
 
     :param arguments: sequence of program arguments (including the program name)
     :type arguments: list[str]
     """
 
     cmd = arguments
     log.debug('Running subprocessio command %s', cmd)
     try:
         _opts = {'shell': False, 'fail_on_stderr': False}
         if env:
             _opts.update({'env': env})
         p = SubprocessIOChunker(cmd, **_opts)
         stdout = ''.join(p)
         stderr = ''.join(''.join(p.error))
     except (EnvironmentError, OSError) as err:
         cmd = ' '.join(cmd)  # human friendly CMD
         tb_err = ("Couldn't run subprocessio command (%s).\n"
                   "Original error was:%s\n" % (cmd, err))
         log.exception(tb_err)
         raise Exception(tb_err)
 
     return stdout, stderr